Changelog

The meaningful changes in Nowledge Mem, written for people using it every day.

Latest release

0.10.30

16 jul 2026

0.10 is a ground-up engine rewrite

The 0.10.x line runs on Mem's rebuilt native engine, so most things are faster and lighter. It is a new engine, though, so you may still hit something unexpected. If you do, open Settings → Info, export your app.log, and send it to hello@nowledge-labs.ai and we'll get on it quickly. And if anything is blocking your work, you can download 0.9.29 from the Downloads page and roll back anytime.

0.10.30

16 jul 2026

0.10.30Latest release16 jul 2026

A Calmer Timeline, Search Across Languages

Memory reviews are easier to work through, search can bridge writing systems, and Library extraction now proves what it saved or linked.

Added

The App interface is now available in Japanese, Korean, Spanish, French, German, and Portuguese, alongside English and Simplified or Traditional Chinese. Dates follow the language you choose, while the language your AI writes in remains an independent Profile setting.
Timeline now gathers pending Memory conflicts and review requests into one focused inbox. The newest decisions remain easy to reach in the App, nmem, or MCP; older untouched items leave the inbox without deleting or changing any Memory.

Improved

Library extraction now completes only when it can show a real outcome: a new source-attributed Memory, a provenance link to knowledge you already had, or a clear explanation that the Source contained nothing worth extracting. Existing Memories are linked without being rewritten.

Fixed

Library now shows completed image-text and PDF Vision work as complete instead of leaving it labeled as queued. Sources carrying an unfinished queue state from an older release become clearly retryable, and opening Library alone never spends managed Vision credits.
Native Gemini providers now receive Gemini-native reasoning and structured-output settings, so an OpenAI-only field can no longer make a valid Gemini request fail before it starts.
Feedback reports now remove credentials and sensitive connection values before attaching diagnostics. On macOS and Windows, Mem can also capture the native App viewport when browser-style screenshots are unavailable.
Memory search can now preserve a strong semantic match across writing systems, such as a Chinese question finding the relevant English Memory, while the existing precision guard still filters unrelated matches in the same language.
The macOS App now recovers when WebKit terminates its content process instead of leaving a blank, frozen window. Permanent backend events are kept outside the WebView, and recovery only returns to a verified local App URL.
Search setup remains usable in compact windows: the dialog stays inside the screen, its actions remain reachable, and longer setup details scroll without moving the header or footer out of view.
APT and RPM upgrades now restart an already-running headless Mem service with the installed version. A service you deliberately stopped stays stopped, and nmem status points out any remaining local CLI and Server version mismatch.
Memory Maintenance previews now stay inside the Space and candidate limit you selected, and rank only active, visible Memories. A small preview no longer looks healthy while eligible work is hidden behind archived or historical rows.
Error notices across the App now offer Copy details, so provider, Skill, connection, and setup failures can be shared without retyping or losing the original message.
Opening Slides now enables the Presentation Creator directly instead of trying to install a Python environment the desktop app no longer ships. Its live preview and PowerPoint export are unchanged.

0.10.29

15 jul 2026

0.10.2915 jul 2026

Safer Labels, Honest Background Work

Clean up labels without putting the knowledge behind them at risk, with background work and Skill verification that report exactly what is current.

Added

You can now merge duplicate labels or remove an unwanted label after reviewing its full impact. Memories, sources, entities, and conversations keep their content; merging moves every assignment to the label you choose, while deleting removes only the label. The same preview-first, replay-safe flow is available in the App, nmem, and writable MCP connections, both locally and in Cloud workspaces.
Memory Policy is now a portable Setting, kept separate from Rules. It records what you want captured into long-term memory and what to skip, travels with your Settings, and never enters an agent's live instructions. Rules still guide how your AI behaves; Memory Policy only shapes what gets remembered.

Improved

Graph maintenance now has one durable job record from start to finish. The App can follow real progress, completed work stays completed after a restart, interrupted work is closed clearly, and a failed provider response no longer leaves a task looking active or successful.
Background AI work now honors the token budget you actually see. A hidden lower cap was removed, the visible budget is enforced while a task runs instead of only at the start, and each task's token use stays accurately attributed in the audit.
You can now hide an event from Timeline to keep it focused, and bring it back whenever you want. The choice is reversible and survives closing and reopening Timeline.

Fixed

Skill verification now belongs to the exact procedure and benchmark that were tested. Editing either one immediately marks the proof as stale across the App, nmem, and MCP, and interrupted or failed checks can no longer appear as a current pass.
Data Transfer now keeps migration history, background-job state, and governance receipts on the installation that owns them. These internal records are excluded from both export and import, so restoring knowledge cannot replay old operations or overwrite the destination's runtime state.

0.10.28

14 jul 2026

0.10.2814 jul 2026

Cleaner Agent Context, Trustworthy Memory History

Mem now keeps what each named AI has learned in the right context, and makes every automatic Memory relationship explanation traceable to the two Memories it describes.

Added

You can now create a Markdown document directly in Library. It stays as an unsaved draft until you choose Save, remains available while you check another source, then becomes searchable and editable like the rest of your Library; importing an existing file remains a separate action.
AI Now can now run reviewed work on a daily, weekly, weekday, selected-day, or monthly schedule. Every run starts with a bounded fresh context and only the access you approved; schedules can be previewed, paused, edited, resumed, deleted, and audited from AI Now or nmem schedules. AI may prepare a draft, but only you can activate it.
Memories, Working Memory, AI Context Bundles, and readable Files now have stable references you can carry between agents and clients. Open a nowledgemem:// reference in the App, read it through MCP Resources, or resolve it with nmem; it always stays inside the local or Cloud workspace you are already connected to and never carries a credential.

Improved

Skill benchmarks can now include realistic near-miss tasks that should not activate a Skill. nmem matching, MCP discovery, and benchmark runs use the same semantic router; if semantic evidence is unavailable, Mem reports the result as inconclusive instead of treating a lexical guess as proof.
Named AI identities can now accumulate their own reviewable working rules from explicitly attributed conversations. A preference learned with one agent stays with that agent, while profiles remain yours to edit and draft rules never enter an agent's context until you approve them.
Profile output language now covers every European Union official language plus Norwegian Bokmål and Icelandic, so more people can have their AI answer in their own language. This is the language the AI writes in, separate from the app's interface language, and your existing choice is unchanged.

Fixed

Access Anywhere no longer keeps restarting Cloudflare's helper when a VPN, proxy, or DNS setup blocks every tunnel path. After a bounded recovery attempt, Settings pauses cleanly, explains whether fake-IP interception, TLS, QUIC, or configuration is blocking the link, and lets you retry deliberately without replacing your API key.
WorkBuddy and CodeBuddy conversation capture now reaches the local history parser instead of being rejected as an unsupported client. Automatic connector sync and nmem history import keep WorkBuddy and CodeBuddy as distinct sources on macOS, Linux, and Windows.
Conversation-history sync now adds one clear Timeline summary after the batch finishes. Imported conversations are searchable immediately without silently queuing one model review per thread; long-term memory extraction remains an explicit, paced follow-up.
Saved Codex conversations now keep their user turns with the current Codex transcript format as well as older versions. Codex and Claude Code capture waits for a real user message instead of saving an incomplete assistant-only snapshot, and safely reconstructs affected captures when possible.
Automatic EVOLVES links and conflict reviews now prove their explanation with an exact excerpt from each Memory. Mem verifies both excerpts before changing the graph or Timeline, so a correct relationship can no longer display a reason copied from an unrelated Memory pair.
Background memory-type review now uses your default text model, leaving the OCR model for image and document work. If the provider is unavailable, Task Center reports a partial run instead of presenting silent fallback classifications as a success.
SuperGrok sign-in now handles the one-time code xAI may show for Grok Build. Browser sign-in still completes automatically when possible; remote and headless users can paste the code into Settings to finish the same secure login.
Custom Responses providers now carry compatibility learned during verification into AI Now and background work. Gateways that require streaming or reject OpenAI's optional output limit no longer pass Settings only to fail on the first real task.
Opening a dated Working Memory snapshot from the Files calendar now resolves the real archive file instead of failing with an unsupported-path error.

0.10.27

13 jul 2026

0.10.2713 jul 2026

Settings That Travel, A Graph You Can Curate

Move the way Mem is configured without moving secrets, and clean up duplicate or unwanted entities without putting the memories behind them at risk.

Added

Settings now has its own backup and restore flow, separate from knowledge Data Transfer. It carries your profile, Spaces, background-intelligence preferences, and provider/model routing to another installation, merging them with settings already there instead of removing destination-only entries. API keys, sign-in sessions, Access Anywhere keys, licenses, and device paths stay on the original device.
You can now merge duplicate entities or remove an unwanted entity from its Knowledge Graph details. Mem shows the impact first: merging keeps the canonical entity and moves its aliases, memory mentions, labels, topics, and custom relationships; deleting removes only the entity and its links, never the memories that mentioned it. The same preview-first controls are available from nmem and full writable MCP connections.

Improved

The Codex connector now loads your current Nowledge context automatically and keeps cross-tool or exact-history questions routed to Nowledge even when Codex's experimental local Memory is enabled. If you use both, the setup guide now shows the one isolation switch that prevents Codex from re-learning MCP and web results into a second, older copy.
Codex setup now follows the version you actually have: current builds no longer receive a removed compatibility switch, older builds still get it when needed, and the guide makes the one-time hook trust review explicit. Context loading and thread capture also handle Windows and WSL command shims, non-English text, and slow startup fallbacks more reliably.
Settings backup now opens the system Save dialog in the desktop App and downloads through the browser on the Web. The completion message names the exact path or filename, so you always know where the portable settings file went.
Knowledge Data Transfer is now a stable backup and restore flow. Settings and knowledge are clearly separated, existing-item behavior is explained before restore, and optional data scope and storage tools stay available without crowding the page.
Selecting an entity in Knowledge Graph now keeps your current Interpret flow in place. Merge or delete is available from the selected entity context and its Inspect panel. The Graph side panel now uses a clear Interpret / Inspect / Maintain flow, with quieter maintenance controls and cleaner scrolling. Timeline relationship events show readable memory titles, both full memories, and a focused text diff instead of exposing internal IDs as content.

Fixed

Capturing an attached Library document from Timeline now saves the knowledge Mem extracted, not the short instruction used to request it. The new Memory stays linked to its source and is included in that document's extracted-memory count. Task Center also names conversation-distillation work by its conversation title across the queue, active work, and recent history.
Search results keep their real source, so memories from Codex, Claude Code, Hermes, Library, and other clients no longer appear as generic Background Intelligence. Knowledge Graph extraction also uses only entities relevant to the current memory and no longer applies ungrounded follow-up rewrites that could change entity names into another language.
A Skill update that was prepared for an older live version is now clearly marked as outdated instead of failing behind an Apply button. You can inspect and discard only that proposal, then prepare a fresh update without changing the Skill your tools currently use.

0.10.26

12 jul 2026

0.10.2612 jul 2026

Ask Your Knowledge, From Anywhere

Questions in Timeline and the terminal now draw from the same evidence-aware answer system across your memories, conversations, Library, and knowledge graph.

Added

The new nmem ask command answers questions from your knowledge in the terminal. Answers stream as they are written and are saved to Timeline by default; --ephemeral keeps one-off questions out of your history, while --json returns structured evidence and search coverage for scripts.

Improved

Knowledge answers now search Memories, past conversations, Library documents, and graph connections together, while respecting each Space's retrieval policy. Timeline shows compact, clickable evidence, and clearly marks an answer as incomplete when part of the knowledge base could not be searched.
Long AI Now conversations keep a compact checkpoint of older work and exact recent turns, reducing repeated model input without shortening the saved task history.

Fixed

Search now heals a stale text index on its own. You still get meaning-based results while it catches up, and the repair runs in small bounded background passes, so a drifted index no longer means missing results or a manual reindex.
OpenClaw conversations sync to Threads again on OpenClaw 2026.6.11. New Threads use the opening request as a readable title instead of exposing an internal session key. The updated plugin also surfaces restricted tool profiles and blocked conversation hooks in status, and keeps the full ten-tool contract available when explicitly granted.
Knowledge Graph now opens without running hidden source-repair work in its render path. Returning to a local Graph reuses the recent topology instead of downloading and laying out the same large graph again; hosted workspaces keep a shorter freshness window for changes from other devices. Large overviews settle in about half as many frames without hiding any nodes or connections.
If an AI Now checkpoint cannot be refreshed, Mem now falls back to the last verified checkpoint plus every following turn, or the complete conversation when no checkpoint is safe. Older context is never silently dropped.
Large background memory-maintenance runs no longer hold the knowledge graph for the entire job. Search, graph views, and health checks remain responsive while freshness scores are refreshed in small bounded batches.

Platform

Hermes Agent now has a native Windows PowerShell installer that uses Hermes' platform directory, validates the complete provider before installation, and preserves any other active memory provider. Git Bash and WSL are no longer required.

0.10.25

11 jul 2026

0.10.2511 jul 2026

Work That Finishes, Data That Stays in Place

This release closes several long-running workflows, adds safer Space and Library controls, and keeps agent work, memories, and conversations consistent when providers or servers are interrupted.

Improved

Spaces can now be consolidated from Settings with a complete preview, and aliases can be managed without the CLI. Library sources can also move between Spaces without losing parsed content, extracted memories, or provenance.
Agent-created memories now check exact and semantic matches before writing, coordinate same-Space updates, and respect a configurable autonomous importance ceiling. Manual and external writes keep the importance you chose.

Fixed

Adding new examples to an existing Skill now prepares a reviewable update instead of appearing to run forever. If the model or server stops before finishing, Mem keeps the examples, shows that the update was interrupted, and lets you retry or keep the current Skill.
Long conversation histories now keep their place while loading more messages, and Jump to latest fetches the true final message window before scrolling. Conversations with thousands of messages no longer stop at the end of the first loaded page.
Recent Skill activity can now identify the connected agent that reported the result, such as Claude Code or Hermes, instead of showing every completion as coming from a generic AI.
AI Now now passes OAuth credentials to connected HTTP MCP plugins in the Rust runtime without writing them into task history. Connected services such as Notion remain available in new, resumed, and reconnected tasks.
Older queued conversation reviews now resume in bounded background waves instead of remaining permanently waiting. Completed work is removed from the backlog, while new conversations still keep their quiet period.
ChatGPT Subscription selections that used the API-only gpt-5.5 model recover to a supported model, and custom Responses providers now consistently send store: false during both connection checks and AI work.
Conversation imports, Space moves, search projections, and cleanup now share one logical conversation identity. Legacy duplicate or empty shells are shown in a review-first cleanup preview and are never removed when they contain messages or provenance.

0.10.24

11 jul 2026

0.10.2411 jul 2026

Reliable Connections and Background Work

Access Anywhere restarts cleanly, background work follows live provider changes, and longer Anthropic-powered agent tasks reuse stable context more efficiently.

Improved

Longer AI Now, Graph Intelligence, Feed, and background-agent runs now reuse stable prompt context when you use Anthropic directly. Mem enables caching only for growing conversations and tool loops, leaves one-shot work uncached, and never sends Anthropic-only fields to Responses, Chat Completions, or unknown custom endpoints.
AI Now's Task List is now built for large histories. Tasks stay in stable, searchable groups such as In progress, Today, and This week; keyboard navigation, a resizable History pane, and one consistent delete confirmation make it faster to move between work without the list shifting underneath you.
Long AI Now runs now keep the answer at the center. Search, reasoning, subagent, and file activity collapses into a concise progress summary with the full trace available on demand, while completed Research evidence and file previews hydrate only near the viewport. Reopening or scrolling through a large task no longer tries to render the entire run at once.

Fixed

Mem now reclaims abandoned Nowledge marketplace staging clones left by interrupted Codex updates, preventing full community repository copies from accumulating indefinitely. Cleanup requires an exact Nowledge Git origin and leaves recent, active, ambiguous, and unrelated marketplace data untouched; new Codex installs continue to use the much smaller sparse checkout.
Data Transfer backups now restore Library files to the destination installation instead of keeping absolute paths from the original machine. Parsed documents open normally after a cross-platform restore, and merge conflicts point to the exact file Mem preserved.
Provider connection tests now give reasoning models enough room to return a short final answer and honor the timeout selected for local or remote endpoints. Mem verifies final text across Chat Completions, Responses, and Anthropic-compatible protocols, and explains when a model returned only reasoning or exhausted its output budget instead of showing a misleading Connected state. During background extraction and distillation, an output-limited reasoning response gets one bounded retry; if it still cannot answer, the task reports the failure and its usage instead of quietly finishing with no result.
Short Memory IDs now work wherever you use them. An eight-character ID copied from a summary, skill, or generated artifact can be opened, updated, linked, expanded on the graph, or passed to an MCP tool directly. Mem resolves it only when the match is unique and asks for more characters instead of guessing when it is not.
You can now enroll every discovered skill at once with `nmem skills enroll --all --yes`. Mem skips unchanged skills that are already managed, completes the remaining imports one by one, and reports any individual failures without losing the successful work. Enroll results also point directly to your drafts, and an incorrect dismiss attempt tells you how to archive a draft instead.
Skills now keep author-defined invocation controls and other YAML settings through import, refinement, editing, and Apply. Refined versions cannot include private benchmark instructions, unsupported `nmem skills` subcommands, or invalid outcome flags, and direct SKILL.md edits can be staged exactly for review. Credential scanning now distinguishes real secret values from documentation, placeholders, and variable names, with safe file-and-line diagnostics when it finds one.
`nmem memories list --limit <n>` now returns the requested number of memories even above the server's safe 100-row page size. The CLI reads bounded pages until it reaches your limit or the end of the list, and JSON output reports both the total available and the number returned.
The `nmem feed` default view now keeps review flags and warning or action-required events visible. When only routine activity exists, it shows the full matching count and points to `--all` instead of reporting an empty Feed. Daily Briefing, Working Memory refresh, and other background tasks also surface provider failures on Timeline and enter the shared rate-limit cooldown instead of failing silently.
Skill Suggestions now finishes once per scheduled cycle instead of reopening another review chain after reaching its safe evidence limit. Each review and continuation also has a real 500K-token safety cap, so accepted suggestions are preserved without repeated background model use.
Access Anywhere no longer leaves hidden Cloudflare tunnel processes behind after an app crash, forced update, or restart during connection setup. Mem reclaims only tunnels owned by the same installation and cancels an in-progress connection promptly when you quit.
Background intelligence now picks up provider changes without restarting Mem. Setting up your first model after launch, switching models, or removing a provider immediately updates distillation, insights, graph extraction, and other scheduled AI work. Custom providers continue to use the selected Responses or Chat Completions protocol.
Local MCP connections now match Mem's real authentication policy. Grok Build and other OAuth-aware clients can connect on the same machine without being sent to an unnecessary sign-in flow; LAN and Access Anywhere endpoints remain protected. Grok Build, Claude Code, Codex, and other recognized agent clients now receive a focused memory toolset through the standard MCP handshake, keeping their model requests lighter without removing the search, save, thread, filesystem, or skill-feedback workflows they need.

0.10.23

10 jul 2026

0.10.2310 jul 2026

More Reliable Research and Provider Connections

This cycle makes long AI Now research tasks more reliable, adds SuperGrok subscription sign-in, and tightens shared-space reads, exports, Windows builds, and multilingual provider setup.

Highlights

AI Now now follows tool execution as a distinct phase of a research turn. A healthy long-running research tool is no longer mistaken for a model stream that stopped responding, while real model and tool stalls still fail with the right diagnosis.
SuperGrok can now power AI Now and agent workflows through xAI OAuth. Sign in from Settings without pasting an API key, and Mem keeps the local subscription credentials refreshed.

Added

Library artifacts can be moved between spaces from the CLI with `nmem library move <source-id> --to-space <space>` or `--all-in-space`. Mem preserves the source id, indexed chunks, and memory provenance links, so you no longer need to delete and re-import a document just to organize it.
SuperGrok is now available as a subscription provider. Sign in with xAI from Settings, and Mem stores and refreshes the local OAuth credentials for AI Now and agent workflows without requiring an API key.
Craft Agent can now connect to Nowledge Mem. Add the Mem source and guide to a Craft workspace for Working Memory, memory search, and durable saves, then import recorded Craft sessions with `nmem t sync --from craft-agent` when you want the real transcript in Threads.
Tencent WorkBuddy and CodeBuddy now have native Nowledge Mem connectors. Install the community marketplace for bundled MCP, startup context, slash commands, and lifecycle hooks that sync local `.workbuddy` or `.codebuddy` transcripts into Mem Threads with `nmem t sync --from workbuddy` or `--from codebuddy`.
Kimi Code support is now native enough for daily use: install the community repo directly from Kimi Code, then the Mem plugin declares its own lifecycle hooks, slash commands, MCP server, and session-start skill, so Kimi Code can sync completed turns, subagent work, compaction boundaries, and interrupted sessions into Mem Threads after `/reload`.

Improved

The Notion plugin now explains its role more clearly in AI Now: it searches your Notion workspace live through MCP, and it does not automatically import Notion pages into Library.

Fixed

AI Now no longer applies the model-stream timeout to a research tool that starts inside the same pending response step. Long tools get their own guarded execution window; real model and tool stalls still stop with the correct diagnosis and preserve completed progress.
Memory, Thread, Library source, and Timeline reads now resolve space names and aliases through the same shared/all retrieval rules as search. Shared results no longer turn into a 404, Timeline filters no longer go empty when the UI sends a display name, and unrelated private spaces remain isolated.
Long Chinese, Japanese, and Korean content no longer risks a UTF-8 boundary crash during truncation, source vision, background processing, skill handling, or thread export. Windows Rust bundles also compile source text explicitly as UTF-8.
Web and remote Memory and Thread exports now keep their raw JSON, Markdown, HTML, and JSONL payloads intact. Export filters reach the backend, browser downloads stay on the active server origin, and generated HTML safely escapes timestamps.
Numbered Kimi Code MCP instances are now recognized during startup and cleanup, so stale duplicate server entries no longer survive because their numeric suffix was parsed as a literal pattern.
Spreadsheet analysis now accepts the same formats as Library import: CSV, TSV, XLSX, XLSM, XLSB, XLS, and ODS. If a workbook imports successfully, AI Now and MCP table tools can inspect its schema and query it instead of rejecting it as unsupported.
Manual Library extraction now has a clear queued state. The API returns `202 Accepted` when extraction is queued or already queued, and `nmem library extract <source-id> --wait` can block until the background task completes.
Search rebuild no longer gets trapped on very large saved-conversation message indexes. When a message store is large enough to hit the risky native full-text rebuild path, Mem now keeps the backend responsive and lets thread search use the best available fallback instead of leaving reindex permanently active.
Large Library files and full search rebuilds now index source chunks in bounded batches and avoid repeatedly merging fresh results through old search-index fragments. This fixes a native write path that could leave ARM64 and headless servers permanently idle after remote embeddings had finished. If a rebuild step fails, Mem restores the previous searchable snapshot instead of leaving a partial index behind.
Fresh Windows installs now start reliably. The installer detects and installs the Microsoft Visual C++ runtime the backend needs, so the app no longer fails to launch on a clean PC. The bundled `nmem` command-line tool is also set up correctly by the app instead of pointing at a runtime that is no longer shipped, and it is removed cleanly on uninstall.
Title bar notifications now stay readable in both light and dark mode. Graph RAM, database recovery, CLI setup, plugin update, and app update popovers no longer show dark text on a dark surface.
Turning off or archiving an active Skill now removes it from connected tools immediately. Mem prunes the active skill copy, re-syncs registered hosts, and records a Timeline event instead of waiting for the next app restart to reconcile the skill list.
Skill duplicate suggestions no longer treat “Not for … use another skill” guidance as evidence that two skills are the same. Those negative routing notes now help agents choose correctly without polluting duplicate detection.
When Mem suggests two active Skills might be duplicates, the backend now has a durable “not a duplicate” dismissal path. Once a pair is dismissed, future merge scans filter it out instead of showing the same pair again.
Manual Memory Knowledge Graph extraction now respects the model path you selected. When the local LLM is installed, on-device extraction no longer falls through to a remote MiniMax-style credential error; choosing Remote still uses your configured provider.
Kimi Code connector guidance now recognizes a stale `nmem` CLI as an outdated client. The bundled skill, slash commands, website guide, and integration registry tell agents to refresh the same CLI source before retrying Kimi-specific MCP or thread-sync commands.
Claude Code and Grok Build no longer guide agents to create project-named spaces from the current folder. Unless you explicitly set `NMEM_SPACE`, the connector reads, searches, and saves in your default Mem space.
`nmem library search-chunks` and `nmem library analyze` now work against the Rust server in local and remote mode. They no longer fail with an MCP initialize-handshake error when you inspect chunks or table data from the CLI.
`nmem import <path>` now recognizes obvious third-party conversation exports such as Claude.ai, Google Takeout, Alma, ChatWise, HTML, and bulk JSON files, then sends them through the conversation importer. A Claude export zip no longer fails as if it were a Mem backup missing `manifest.json`.
`nmem memories add --id <existing-id>` works as an upsert again. Scripts that save with stable IDs now update the existing memory, preserve its graph relationships, and merge metadata instead of failing with a duplicate-primary-key storage error.
Space details now show Working Memory usage from the same check that protects deletion. A space with current Working Memory no longer looks empty until the delete action suddenly asks you to purge it.
`nmem spaces delete --dry-run` now previews every blocker before an irreversible space removal. It lists memory, thread, source, and Working Memory presence, and only suggests `--purge-working-memory` when Working Memory is the remaining blocker.
Space creation errors now identify the conflicting field and owning space when an `id` or `key` matches an existing space name or alias. The error no longer blames the display name you were trying to create when the real conflict is a retained alias.
Space aliases are now visible in `nmem spaces list` and manageable from `nmem spaces update` with `--alias`, `--remove-alias`, and `--clear-aliases`. Alias updates no longer return success while silently leaving retained aliases in place.
Moving specific memories or threads by ID now uses each item's real current space when you do not pass `--space`. Moving a single item back to Default no longer fails just because your CLI is already focused on Default.

0.10.22

9 jul 2026

0.10.229 jul 2026

Sharper Table Analysis and Document Import

AI Now can now answer more precise questions about CSV and Excel Library files by inspecting the table schema first, then running safe read-only SQL for exact aggregations, explicit denominators, and chart-ready results. Library import also reads Office and PDF documents more faithfully, keeping original files intact while adding cleaner parsed text and visual context when a Vision model is available.

Highlights

Ask AI Now for counts, breakdowns, grouped totals, or chart-ready summaries from CSV and Excel files. Mem now reads the table structure first, then runs a lightweight read-only SQL query over the selected sheet instead of relying only on broad table profiling.

Added

AI Now includes schema-first table query tools for tabular Library sources and artifacts. Queries are read-only, capped, and return small structured result sets that can feed the existing chart renderer.

Improved

AI Now now gathers related table metrics in one pass, keeps the chat quieter while it works, and prefers readable horizontal bar charts when category labels are long.
PDF import now separates pages that need OCR from text-native pages that need visual understanding for charts, screenshots, figures, or tables. Mem also handles crop boxes, lists, tables, spacing, and subscripts more cleanly before any Vision pass runs.

Fixed

Word, PowerPoint, and Excel Library imports now keep the original Office file separate from the parsed Markdown. Preview, reindex, and table-analysis paths continue to use the real `.docx`, `.pptx`, `.xlsx`, or `.xls` file instead of a Markdown file saved under the original extension.
Word documents with tables no longer expose internal OOXML formatting tags in parsed text. The table content stays searchable without showing border and layout markup to the user.
Background Processing now respects the hourly AI limit's unlimited setting even when a headless or remote server receives it through nested settings responses. Turning the hourly limit off no longer leaves work queued behind a local hourly batch cap.
Background intelligence now keeps up with what you save. The work a saved or imported conversation kicks off (creating memories, extracting the knowledge graph, refreshing Working Memory, follow-up review) is no longer spread thin across an hourly background task limit, so a freshly saved thread is processed promptly instead of looking stuck. Only periodic always-on maintenance stays gently paced, and your total AI usage is still bounded by your hourly limit.
When you run a Nowledge Mem skill from AI Now, Mem now includes that exact skill procedure in the turn. A selected skill no longer gets lost when the enabled-skill prompt budget is tight.
AI Now plugins are now available to the current task as soon as you connect or configure them. Obsidian vaults, Notion OAuth connections, API-key plugins, and custom MCP connectors no longer look ready in Settings while the agent still cannot see their tools.
AI Now research tasks now recover more cleanly when the selected model stops streaming after visible progress. If tools, plans, or permissions have already appeared, Mem keeps that completed context and asks you to Continue from the transcript instead of replaying the same turn and risking repeated tool work.

0.10.21

8 jul 2026

0.10.218 jul 2026

Excel Analysis and Startup Reliability

This hotfix restores computed analysis for Excel files, makes rare desktop blank-window failures easier to recover from and diagnose, and brings the browser extension back to Firefox for web AI chat capture.

Highlights

Background insights and crystals now run on a calm schedule and finish cleanly. Previously, on an active library, they could restart on nearly every background check, use a lot of AI, and fill the background task list with runs marked as paused. Each pass now completes once for its cycle, keeps what it produced, and shows as completed.
AI Now and MCP can now compute column statistics, categorical summaries, date ranges, and cross-column patterns from Excel `.xlsx` and `.xls` Library files instead of treating them like empty tables.
The desktop app now takes a safer local live-update path and shows a recovery screen with a logged cause if the frontend crashes, instead of leaving a silent blank window.
Nowledge Mem Exchange is available for Firefox again. Firefox uses its native sidebar for capture, Manual Distill, and Thread Backup; Browse Now browser control still requires Chrome, Edge, Arc, or another Chromium browser.

Fixed

`analyze_source_data` and `analyze_artifact_data` now read real workbook sheets before running tabular analysis. CSV and TSV behavior is unchanged, and unreadable workbooks now return a clear error instead of a misleading empty-table result.
Local desktop live updates now use the browser-native EventSource path when loopback auth is not required. This avoids a WebKit failure mode where the backend keeps running but the main window can go blank a few seconds after startup.
Frontend runtime crashes now write a compact reason to `app.log` and show a reload recovery screen. Successful Browser Extension WebSocket upgrades also no longer appear as backend non-2xx warnings.
Graph Intelligence and AI Now recover from short model-provider connection blips before any answer has appeared. Mem retries those safe failures automatically, while rate limits and quota errors still use the normal cooldown path.
Knowledge Graph search, resize, and initial canvas render now handle asynchronous layout failures gracefully, so a graph that loaded correctly is less likely to be interrupted by a G6 layout race.
Data Transfer now includes Timeline history. Restored or cloud-imported workspaces keep past briefing, review, source-ingest, and maintenance cards instead of arriving with an empty Timeline.
AI Now keeps skill prompt injection bounded so several enabled skills or a large custom skill are less likely to overwhelm local or OpenAI-compatible models. If a selected model still stops streaming, Mem now points users to the model/provider path instead of the Nowledge service status page.

Platform

Ready to try Mem?

Download the app, connect your AI tools, and keep useful context where you control it.

Download Read Docs

Changelog

The meaningful changes in Nowledge Mem, written for people using it every day.

Latest release

0.10.30

16 jul 2026

0.10 is a ground-up engine rewrite

0.10.30

16 jul 2026

0.10.30Latest release16 jul 2026

A Calmer Timeline, Search Across Languages

Memory reviews are easier to work through, search can bridge writing systems, and Library extraction now proves what it saved or linked.

Added

The App interface is now available in Japanese, Korean, Spanish, French, German, and Portuguese, alongside English and Simplified or Traditional Chinese. Dates follow the language you choose, while the language your AI writes in remains an independent Profile setting.
Timeline now gathers pending Memory conflicts and review requests into one focused inbox. The newest decisions remain easy to reach in the App, nmem, or MCP; older untouched items leave the inbox without deleting or changing any Memory.

Improved

Library extraction now completes only when it can show a real outcome: a new source-attributed Memory, a provenance link to knowledge you already had, or a clear explanation that the Source contained nothing worth extracting. Existing Memories are linked without being rewritten.

Fixed

Library now shows completed image-text and PDF Vision work as complete instead of leaving it labeled as queued. Sources carrying an unfinished queue state from an older release become clearly retryable, and opening Library alone never spends managed Vision credits.
Native Gemini providers now receive Gemini-native reasoning and structured-output settings, so an OpenAI-only field can no longer make a valid Gemini request fail before it starts.
Feedback reports now remove credentials and sensitive connection values before attaching diagnostics. On macOS and Windows, Mem can also capture the native App viewport when browser-style screenshots are unavailable.
Memory search can now preserve a strong semantic match across writing systems, such as a Chinese question finding the relevant English Memory, while the existing precision guard still filters unrelated matches in the same language.
The macOS App now recovers when WebKit terminates its content process instead of leaving a blank, frozen window. Permanent backend events are kept outside the WebView, and recovery only returns to a verified local App URL.
Search setup remains usable in compact windows: the dialog stays inside the screen, its actions remain reachable, and longer setup details scroll without moving the header or footer out of view.
APT and RPM upgrades now restart an already-running headless Mem service with the installed version. A service you deliberately stopped stays stopped, and nmem status points out any remaining local CLI and Server version mismatch.
Memory Maintenance previews now stay inside the Space and candidate limit you selected, and rank only active, visible Memories. A small preview no longer looks healthy while eligible work is hidden behind archived or historical rows.
Error notices across the App now offer Copy details, so provider, Skill, connection, and setup failures can be shared without retyping or losing the original message.
Opening Slides now enables the Presentation Creator directly instead of trying to install a Python environment the desktop app no longer ships. Its live preview and PowerPoint export are unchanged.

0.10.29

15 jul 2026

0.10.2915 jul 2026

Safer Labels, Honest Background Work

Clean up labels without putting the knowledge behind them at risk, with background work and Skill verification that report exactly what is current.

Added

You can now merge duplicate labels or remove an unwanted label after reviewing its full impact. Memories, sources, entities, and conversations keep their content; merging moves every assignment to the label you choose, while deleting removes only the label. The same preview-first, replay-safe flow is available in the App, nmem, and writable MCP connections, both locally and in Cloud workspaces.
Memory Policy is now a portable Setting, kept separate from Rules. It records what you want captured into long-term memory and what to skip, travels with your Settings, and never enters an agent's live instructions. Rules still guide how your AI behaves; Memory Policy only shapes what gets remembered.

Improved

Graph maintenance now has one durable job record from start to finish. The App can follow real progress, completed work stays completed after a restart, interrupted work is closed clearly, and a failed provider response no longer leaves a task looking active or successful.
Background AI work now honors the token budget you actually see. A hidden lower cap was removed, the visible budget is enforced while a task runs instead of only at the start, and each task's token use stays accurately attributed in the audit.
You can now hide an event from Timeline to keep it focused, and bring it back whenever you want. The choice is reversible and survives closing and reopening Timeline.

Fixed

Skill verification now belongs to the exact procedure and benchmark that were tested. Editing either one immediately marks the proof as stale across the App, nmem, and MCP, and interrupted or failed checks can no longer appear as a current pass.
Data Transfer now keeps migration history, background-job state, and governance receipts on the installation that owns them. These internal records are excluded from both export and import, so restoring knowledge cannot replay old operations or overwrite the destination's runtime state.

0.10.28

14 jul 2026

0.10.2814 jul 2026

Cleaner Agent Context, Trustworthy Memory History

Mem now keeps what each named AI has learned in the right context, and makes every automatic Memory relationship explanation traceable to the two Memories it describes.

Added

You can now create a Markdown document directly in Library. It stays as an unsaved draft until you choose Save, remains available while you check another source, then becomes searchable and editable like the rest of your Library; importing an existing file remains a separate action.
AI Now can now run reviewed work on a daily, weekly, weekday, selected-day, or monthly schedule. Every run starts with a bounded fresh context and only the access you approved; schedules can be previewed, paused, edited, resumed, deleted, and audited from AI Now or nmem schedules. AI may prepare a draft, but only you can activate it.
Memories, Working Memory, AI Context Bundles, and readable Files now have stable references you can carry between agents and clients. Open a nowledgemem:// reference in the App, read it through MCP Resources, or resolve it with nmem; it always stays inside the local or Cloud workspace you are already connected to and never carries a credential.

Improved

Skill benchmarks can now include realistic near-miss tasks that should not activate a Skill. nmem matching, MCP discovery, and benchmark runs use the same semantic router; if semantic evidence is unavailable, Mem reports the result as inconclusive instead of treating a lexical guess as proof.
Named AI identities can now accumulate their own reviewable working rules from explicitly attributed conversations. A preference learned with one agent stays with that agent, while profiles remain yours to edit and draft rules never enter an agent's context until you approve them.
Profile output language now covers every European Union official language plus Norwegian Bokmål and Icelandic, so more people can have their AI answer in their own language. This is the language the AI writes in, separate from the app's interface language, and your existing choice is unchanged.

Fixed

Access Anywhere no longer keeps restarting Cloudflare's helper when a VPN, proxy, or DNS setup blocks every tunnel path. After a bounded recovery attempt, Settings pauses cleanly, explains whether fake-IP interception, TLS, QUIC, or configuration is blocking the link, and lets you retry deliberately without replacing your API key.
WorkBuddy and CodeBuddy conversation capture now reaches the local history parser instead of being rejected as an unsupported client. Automatic connector sync and nmem history import keep WorkBuddy and CodeBuddy as distinct sources on macOS, Linux, and Windows.
Conversation-history sync now adds one clear Timeline summary after the batch finishes. Imported conversations are searchable immediately without silently queuing one model review per thread; long-term memory extraction remains an explicit, paced follow-up.
Saved Codex conversations now keep their user turns with the current Codex transcript format as well as older versions. Codex and Claude Code capture waits for a real user message instead of saving an incomplete assistant-only snapshot, and safely reconstructs affected captures when possible.
Automatic EVOLVES links and conflict reviews now prove their explanation with an exact excerpt from each Memory. Mem verifies both excerpts before changing the graph or Timeline, so a correct relationship can no longer display a reason copied from an unrelated Memory pair.
Background memory-type review now uses your default text model, leaving the OCR model for image and document work. If the provider is unavailable, Task Center reports a partial run instead of presenting silent fallback classifications as a success.
SuperGrok sign-in now handles the one-time code xAI may show for Grok Build. Browser sign-in still completes automatically when possible; remote and headless users can paste the code into Settings to finish the same secure login.
Custom Responses providers now carry compatibility learned during verification into AI Now and background work. Gateways that require streaming or reject OpenAI's optional output limit no longer pass Settings only to fail on the first real task.
Opening a dated Working Memory snapshot from the Files calendar now resolves the real archive file instead of failing with an unsupported-path error.

0.10.27

13 jul 2026

0.10.2713 jul 2026

Settings That Travel, A Graph You Can Curate

Move the way Mem is configured without moving secrets, and clean up duplicate or unwanted entities without putting the memories behind them at risk.

Added

Settings now has its own backup and restore flow, separate from knowledge Data Transfer. It carries your profile, Spaces, background-intelligence preferences, and provider/model routing to another installation, merging them with settings already there instead of removing destination-only entries. API keys, sign-in sessions, Access Anywhere keys, licenses, and device paths stay on the original device.
You can now merge duplicate entities or remove an unwanted entity from its Knowledge Graph details. Mem shows the impact first: merging keeps the canonical entity and moves its aliases, memory mentions, labels, topics, and custom relationships; deleting removes only the entity and its links, never the memories that mentioned it. The same preview-first controls are available from nmem and full writable MCP connections.

Improved

The Codex connector now loads your current Nowledge context automatically and keeps cross-tool or exact-history questions routed to Nowledge even when Codex's experimental local Memory is enabled. If you use both, the setup guide now shows the one isolation switch that prevents Codex from re-learning MCP and web results into a second, older copy.
Codex setup now follows the version you actually have: current builds no longer receive a removed compatibility switch, older builds still get it when needed, and the guide makes the one-time hook trust review explicit. Context loading and thread capture also handle Windows and WSL command shims, non-English text, and slow startup fallbacks more reliably.
Settings backup now opens the system Save dialog in the desktop App and downloads through the browser on the Web. The completion message names the exact path or filename, so you always know where the portable settings file went.
Knowledge Data Transfer is now a stable backup and restore flow. Settings and knowledge are clearly separated, existing-item behavior is explained before restore, and optional data scope and storage tools stay available without crowding the page.
Selecting an entity in Knowledge Graph now keeps your current Interpret flow in place. Merge or delete is available from the selected entity context and its Inspect panel. The Graph side panel now uses a clear Interpret / Inspect / Maintain flow, with quieter maintenance controls and cleaner scrolling. Timeline relationship events show readable memory titles, both full memories, and a focused text diff instead of exposing internal IDs as content.

Fixed

Capturing an attached Library document from Timeline now saves the knowledge Mem extracted, not the short instruction used to request it. The new Memory stays linked to its source and is included in that document's extracted-memory count. Task Center also names conversation-distillation work by its conversation title across the queue, active work, and recent history.
Search results keep their real source, so memories from Codex, Claude Code, Hermes, Library, and other clients no longer appear as generic Background Intelligence. Knowledge Graph extraction also uses only entities relevant to the current memory and no longer applies ungrounded follow-up rewrites that could change entity names into another language.
A Skill update that was prepared for an older live version is now clearly marked as outdated instead of failing behind an Apply button. You can inspect and discard only that proposal, then prepare a fresh update without changing the Skill your tools currently use.

0.10.26

12 jul 2026

0.10.2612 jul 2026

Ask Your Knowledge, From Anywhere

Questions in Timeline and the terminal now draw from the same evidence-aware answer system across your memories, conversations, Library, and knowledge graph.

Added

The new nmem ask command answers questions from your knowledge in the terminal. Answers stream as they are written and are saved to Timeline by default; --ephemeral keeps one-off questions out of your history, while --json returns structured evidence and search coverage for scripts.

Improved

Knowledge answers now search Memories, past conversations, Library documents, and graph connections together, while respecting each Space's retrieval policy. Timeline shows compact, clickable evidence, and clearly marks an answer as incomplete when part of the knowledge base could not be searched.
Long AI Now conversations keep a compact checkpoint of older work and exact recent turns, reducing repeated model input without shortening the saved task history.

Fixed

Search now heals a stale text index on its own. You still get meaning-based results while it catches up, and the repair runs in small bounded background passes, so a drifted index no longer means missing results or a manual reindex.
OpenClaw conversations sync to Threads again on OpenClaw 2026.6.11. New Threads use the opening request as a readable title instead of exposing an internal session key. The updated plugin also surfaces restricted tool profiles and blocked conversation hooks in status, and keeps the full ten-tool contract available when explicitly granted.
Knowledge Graph now opens without running hidden source-repair work in its render path. Returning to a local Graph reuses the recent topology instead of downloading and laying out the same large graph again; hosted workspaces keep a shorter freshness window for changes from other devices. Large overviews settle in about half as many frames without hiding any nodes or connections.
If an AI Now checkpoint cannot be refreshed, Mem now falls back to the last verified checkpoint plus every following turn, or the complete conversation when no checkpoint is safe. Older context is never silently dropped.
Large background memory-maintenance runs no longer hold the knowledge graph for the entire job. Search, graph views, and health checks remain responsive while freshness scores are refreshed in small bounded batches.

Platform

Hermes Agent now has a native Windows PowerShell installer that uses Hermes' platform directory, validates the complete provider before installation, and preserves any other active memory provider. Git Bash and WSL are no longer required.

0.10.25

11 jul 2026

0.10.2511 jul 2026

Work That Finishes, Data That Stays in Place

This release closes several long-running workflows, adds safer Space and Library controls, and keeps agent work, memories, and conversations consistent when providers or servers are interrupted.

Improved

Spaces can now be consolidated from Settings with a complete preview, and aliases can be managed without the CLI. Library sources can also move between Spaces without losing parsed content, extracted memories, or provenance.
Agent-created memories now check exact and semantic matches before writing, coordinate same-Space updates, and respect a configurable autonomous importance ceiling. Manual and external writes keep the importance you chose.

Fixed

Adding new examples to an existing Skill now prepares a reviewable update instead of appearing to run forever. If the model or server stops before finishing, Mem keeps the examples, shows that the update was interrupted, and lets you retry or keep the current Skill.
Long conversation histories now keep their place while loading more messages, and Jump to latest fetches the true final message window before scrolling. Conversations with thousands of messages no longer stop at the end of the first loaded page.
Recent Skill activity can now identify the connected agent that reported the result, such as Claude Code or Hermes, instead of showing every completion as coming from a generic AI.
AI Now now passes OAuth credentials to connected HTTP MCP plugins in the Rust runtime without writing them into task history. Connected services such as Notion remain available in new, resumed, and reconnected tasks.
Older queued conversation reviews now resume in bounded background waves instead of remaining permanently waiting. Completed work is removed from the backlog, while new conversations still keep their quiet period.
ChatGPT Subscription selections that used the API-only gpt-5.5 model recover to a supported model, and custom Responses providers now consistently send store: false during both connection checks and AI work.
Conversation imports, Space moves, search projections, and cleanup now share one logical conversation identity. Legacy duplicate or empty shells are shown in a review-first cleanup preview and are never removed when they contain messages or provenance.

0.10.24

11 jul 2026

0.10.2411 jul 2026

Reliable Connections and Background Work

Access Anywhere restarts cleanly, background work follows live provider changes, and longer Anthropic-powered agent tasks reuse stable context more efficiently.

Improved

Longer AI Now, Graph Intelligence, Feed, and background-agent runs now reuse stable prompt context when you use Anthropic directly. Mem enables caching only for growing conversations and tool loops, leaves one-shot work uncached, and never sends Anthropic-only fields to Responses, Chat Completions, or unknown custom endpoints.
AI Now's Task List is now built for large histories. Tasks stay in stable, searchable groups such as In progress, Today, and This week; keyboard navigation, a resizable History pane, and one consistent delete confirmation make it faster to move between work without the list shifting underneath you.
Long AI Now runs now keep the answer at the center. Search, reasoning, subagent, and file activity collapses into a concise progress summary with the full trace available on demand, while completed Research evidence and file previews hydrate only near the viewport. Reopening or scrolling through a large task no longer tries to render the entire run at once.

Fixed

Mem now reclaims abandoned Nowledge marketplace staging clones left by interrupted Codex updates, preventing full community repository copies from accumulating indefinitely. Cleanup requires an exact Nowledge Git origin and leaves recent, active, ambiguous, and unrelated marketplace data untouched; new Codex installs continue to use the much smaller sparse checkout.
Data Transfer backups now restore Library files to the destination installation instead of keeping absolute paths from the original machine. Parsed documents open normally after a cross-platform restore, and merge conflicts point to the exact file Mem preserved.
Provider connection tests now give reasoning models enough room to return a short final answer and honor the timeout selected for local or remote endpoints. Mem verifies final text across Chat Completions, Responses, and Anthropic-compatible protocols, and explains when a model returned only reasoning or exhausted its output budget instead of showing a misleading Connected state. During background extraction and distillation, an output-limited reasoning response gets one bounded retry; if it still cannot answer, the task reports the failure and its usage instead of quietly finishing with no result.
Short Memory IDs now work wherever you use them. An eight-character ID copied from a summary, skill, or generated artifact can be opened, updated, linked, expanded on the graph, or passed to an MCP tool directly. Mem resolves it only when the match is unique and asks for more characters instead of guessing when it is not.
You can now enroll every discovered skill at once with `nmem skills enroll --all --yes`. Mem skips unchanged skills that are already managed, completes the remaining imports one by one, and reports any individual failures without losing the successful work. Enroll results also point directly to your drafts, and an incorrect dismiss attempt tells you how to archive a draft instead.
Skills now keep author-defined invocation controls and other YAML settings through import, refinement, editing, and Apply. Refined versions cannot include private benchmark instructions, unsupported `nmem skills` subcommands, or invalid outcome flags, and direct SKILL.md edits can be staged exactly for review. Credential scanning now distinguishes real secret values from documentation, placeholders, and variable names, with safe file-and-line diagnostics when it finds one.
`nmem memories list --limit <n>` now returns the requested number of memories even above the server's safe 100-row page size. The CLI reads bounded pages until it reaches your limit or the end of the list, and JSON output reports both the total available and the number returned.
The `nmem feed` default view now keeps review flags and warning or action-required events visible. When only routine activity exists, it shows the full matching count and points to `--all` instead of reporting an empty Feed. Daily Briefing, Working Memory refresh, and other background tasks also surface provider failures on Timeline and enter the shared rate-limit cooldown instead of failing silently.
Skill Suggestions now finishes once per scheduled cycle instead of reopening another review chain after reaching its safe evidence limit. Each review and continuation also has a real 500K-token safety cap, so accepted suggestions are preserved without repeated background model use.
Access Anywhere no longer leaves hidden Cloudflare tunnel processes behind after an app crash, forced update, or restart during connection setup. Mem reclaims only tunnels owned by the same installation and cancels an in-progress connection promptly when you quit.
Background intelligence now picks up provider changes without restarting Mem. Setting up your first model after launch, switching models, or removing a provider immediately updates distillation, insights, graph extraction, and other scheduled AI work. Custom providers continue to use the selected Responses or Chat Completions protocol.
Local MCP connections now match Mem's real authentication policy. Grok Build and other OAuth-aware clients can connect on the same machine without being sent to an unnecessary sign-in flow; LAN and Access Anywhere endpoints remain protected. Grok Build, Claude Code, Codex, and other recognized agent clients now receive a focused memory toolset through the standard MCP handshake, keeping their model requests lighter without removing the search, save, thread, filesystem, or skill-feedback workflows they need.

0.10.23

10 jul 2026

0.10.2310 jul 2026

More Reliable Research and Provider Connections

This cycle makes long AI Now research tasks more reliable, adds SuperGrok subscription sign-in, and tightens shared-space reads, exports, Windows builds, and multilingual provider setup.

Highlights

AI Now now follows tool execution as a distinct phase of a research turn. A healthy long-running research tool is no longer mistaken for a model stream that stopped responding, while real model and tool stalls still fail with the right diagnosis.
SuperGrok can now power AI Now and agent workflows through xAI OAuth. Sign in from Settings without pasting an API key, and Mem keeps the local subscription credentials refreshed.

Added

Library artifacts can be moved between spaces from the CLI with `nmem library move <source-id> --to-space <space>` or `--all-in-space`. Mem preserves the source id, indexed chunks, and memory provenance links, so you no longer need to delete and re-import a document just to organize it.
SuperGrok is now available as a subscription provider. Sign in with xAI from Settings, and Mem stores and refreshes the local OAuth credentials for AI Now and agent workflows without requiring an API key.
Craft Agent can now connect to Nowledge Mem. Add the Mem source and guide to a Craft workspace for Working Memory, memory search, and durable saves, then import recorded Craft sessions with `nmem t sync --from craft-agent` when you want the real transcript in Threads.
Tencent WorkBuddy and CodeBuddy now have native Nowledge Mem connectors. Install the community marketplace for bundled MCP, startup context, slash commands, and lifecycle hooks that sync local `.workbuddy` or `.codebuddy` transcripts into Mem Threads with `nmem t sync --from workbuddy` or `--from codebuddy`.
Kimi Code support is now native enough for daily use: install the community repo directly from Kimi Code, then the Mem plugin declares its own lifecycle hooks, slash commands, MCP server, and session-start skill, so Kimi Code can sync completed turns, subagent work, compaction boundaries, and interrupted sessions into Mem Threads after `/reload`.

Improved

The Notion plugin now explains its role more clearly in AI Now: it searches your Notion workspace live through MCP, and it does not automatically import Notion pages into Library.

Fixed

AI Now no longer applies the model-stream timeout to a research tool that starts inside the same pending response step. Long tools get their own guarded execution window; real model and tool stalls still stop with the correct diagnosis and preserve completed progress.
Memory, Thread, Library source, and Timeline reads now resolve space names and aliases through the same shared/all retrieval rules as search. Shared results no longer turn into a 404, Timeline filters no longer go empty when the UI sends a display name, and unrelated private spaces remain isolated.
Long Chinese, Japanese, and Korean content no longer risks a UTF-8 boundary crash during truncation, source vision, background processing, skill handling, or thread export. Windows Rust bundles also compile source text explicitly as UTF-8.
Web and remote Memory and Thread exports now keep their raw JSON, Markdown, HTML, and JSONL payloads intact. Export filters reach the backend, browser downloads stay on the active server origin, and generated HTML safely escapes timestamps.
Numbered Kimi Code MCP instances are now recognized during startup and cleanup, so stale duplicate server entries no longer survive because their numeric suffix was parsed as a literal pattern.
Spreadsheet analysis now accepts the same formats as Library import: CSV, TSV, XLSX, XLSM, XLSB, XLS, and ODS. If a workbook imports successfully, AI Now and MCP table tools can inspect its schema and query it instead of rejecting it as unsupported.
Manual Library extraction now has a clear queued state. The API returns `202 Accepted` when extraction is queued or already queued, and `nmem library extract <source-id> --wait` can block until the background task completes.
Search rebuild no longer gets trapped on very large saved-conversation message indexes. When a message store is large enough to hit the risky native full-text rebuild path, Mem now keeps the backend responsive and lets thread search use the best available fallback instead of leaving reindex permanently active.
Large Library files and full search rebuilds now index source chunks in bounded batches and avoid repeatedly merging fresh results through old search-index fragments. This fixes a native write path that could leave ARM64 and headless servers permanently idle after remote embeddings had finished. If a rebuild step fails, Mem restores the previous searchable snapshot instead of leaving a partial index behind.
Fresh Windows installs now start reliably. The installer detects and installs the Microsoft Visual C++ runtime the backend needs, so the app no longer fails to launch on a clean PC. The bundled `nmem` command-line tool is also set up correctly by the app instead of pointing at a runtime that is no longer shipped, and it is removed cleanly on uninstall.
Title bar notifications now stay readable in both light and dark mode. Graph RAM, database recovery, CLI setup, plugin update, and app update popovers no longer show dark text on a dark surface.
Turning off or archiving an active Skill now removes it from connected tools immediately. Mem prunes the active skill copy, re-syncs registered hosts, and records a Timeline event instead of waiting for the next app restart to reconcile the skill list.
Skill duplicate suggestions no longer treat “Not for … use another skill” guidance as evidence that two skills are the same. Those negative routing notes now help agents choose correctly without polluting duplicate detection.
When Mem suggests two active Skills might be duplicates, the backend now has a durable “not a duplicate” dismissal path. Once a pair is dismissed, future merge scans filter it out instead of showing the same pair again.
Manual Memory Knowledge Graph extraction now respects the model path you selected. When the local LLM is installed, on-device extraction no longer falls through to a remote MiniMax-style credential error; choosing Remote still uses your configured provider.
Kimi Code connector guidance now recognizes a stale `nmem` CLI as an outdated client. The bundled skill, slash commands, website guide, and integration registry tell agents to refresh the same CLI source before retrying Kimi-specific MCP or thread-sync commands.
Claude Code and Grok Build no longer guide agents to create project-named spaces from the current folder. Unless you explicitly set `NMEM_SPACE`, the connector reads, searches, and saves in your default Mem space.
`nmem library search-chunks` and `nmem library analyze` now work against the Rust server in local and remote mode. They no longer fail with an MCP initialize-handshake error when you inspect chunks or table data from the CLI.
`nmem import <path>` now recognizes obvious third-party conversation exports such as Claude.ai, Google Takeout, Alma, ChatWise, HTML, and bulk JSON files, then sends them through the conversation importer. A Claude export zip no longer fails as if it were a Mem backup missing `manifest.json`.
`nmem memories add --id <existing-id>` works as an upsert again. Scripts that save with stable IDs now update the existing memory, preserve its graph relationships, and merge metadata instead of failing with a duplicate-primary-key storage error.
Space details now show Working Memory usage from the same check that protects deletion. A space with current Working Memory no longer looks empty until the delete action suddenly asks you to purge it.
`nmem spaces delete --dry-run` now previews every blocker before an irreversible space removal. It lists memory, thread, source, and Working Memory presence, and only suggests `--purge-working-memory` when Working Memory is the remaining blocker.
Space creation errors now identify the conflicting field and owning space when an `id` or `key` matches an existing space name or alias. The error no longer blames the display name you were trying to create when the real conflict is a retained alias.
Space aliases are now visible in `nmem spaces list` and manageable from `nmem spaces update` with `--alias`, `--remove-alias`, and `--clear-aliases`. Alias updates no longer return success while silently leaving retained aliases in place.
Moving specific memories or threads by ID now uses each item's real current space when you do not pass `--space`. Moving a single item back to Default no longer fails just because your CLI is already focused on Default.

0.10.22

9 jul 2026

0.10.229 jul 2026

Sharper Table Analysis and Document Import

Highlights

Ask AI Now for counts, breakdowns, grouped totals, or chart-ready summaries from CSV and Excel files. Mem now reads the table structure first, then runs a lightweight read-only SQL query over the selected sheet instead of relying only on broad table profiling.

Added

AI Now includes schema-first table query tools for tabular Library sources and artifacts. Queries are read-only, capped, and return small structured result sets that can feed the existing chart renderer.

Improved

AI Now now gathers related table metrics in one pass, keeps the chat quieter while it works, and prefers readable horizontal bar charts when category labels are long.
PDF import now separates pages that need OCR from text-native pages that need visual understanding for charts, screenshots, figures, or tables. Mem also handles crop boxes, lists, tables, spacing, and subscripts more cleanly before any Vision pass runs.

Fixed

Word, PowerPoint, and Excel Library imports now keep the original Office file separate from the parsed Markdown. Preview, reindex, and table-analysis paths continue to use the real `.docx`, `.pptx`, `.xlsx`, or `.xls` file instead of a Markdown file saved under the original extension.
Word documents with tables no longer expose internal OOXML formatting tags in parsed text. The table content stays searchable without showing border and layout markup to the user.
Background Processing now respects the hourly AI limit's unlimited setting even when a headless or remote server receives it through nested settings responses. Turning the hourly limit off no longer leaves work queued behind a local hourly batch cap.
Background intelligence now keeps up with what you save. The work a saved or imported conversation kicks off (creating memories, extracting the knowledge graph, refreshing Working Memory, follow-up review) is no longer spread thin across an hourly background task limit, so a freshly saved thread is processed promptly instead of looking stuck. Only periodic always-on maintenance stays gently paced, and your total AI usage is still bounded by your hourly limit.
When you run a Nowledge Mem skill from AI Now, Mem now includes that exact skill procedure in the turn. A selected skill no longer gets lost when the enabled-skill prompt budget is tight.
AI Now plugins are now available to the current task as soon as you connect or configure them. Obsidian vaults, Notion OAuth connections, API-key plugins, and custom MCP connectors no longer look ready in Settings while the agent still cannot see their tools.
AI Now research tasks now recover more cleanly when the selected model stops streaming after visible progress. If tools, plans, or permissions have already appeared, Mem keeps that completed context and asks you to Continue from the transcript instead of replaying the same turn and risking repeated tool work.

0.10.21

8 jul 2026

0.10.218 jul 2026

Excel Analysis and Startup Reliability

Highlights

Background insights and crystals now run on a calm schedule and finish cleanly. Previously, on an active library, they could restart on nearly every background check, use a lot of AI, and fill the background task list with runs marked as paused. Each pass now completes once for its cycle, keeps what it produced, and shows as completed.
AI Now and MCP can now compute column statistics, categorical summaries, date ranges, and cross-column patterns from Excel `.xlsx` and `.xls` Library files instead of treating them like empty tables.
The desktop app now takes a safer local live-update path and shows a recovery screen with a logged cause if the frontend crashes, instead of leaving a silent blank window.
Nowledge Mem Exchange is available for Firefox again. Firefox uses its native sidebar for capture, Manual Distill, and Thread Backup; Browse Now browser control still requires Chrome, Edge, Arc, or another Chromium browser.

Fixed

`analyze_source_data` and `analyze_artifact_data` now read real workbook sheets before running tabular analysis. CSV and TSV behavior is unchanged, and unreadable workbooks now return a clear error instead of a misleading empty-table result.
Local desktop live updates now use the browser-native EventSource path when loopback auth is not required. This avoids a WebKit failure mode where the backend keeps running but the main window can go blank a few seconds after startup.
Frontend runtime crashes now write a compact reason to `app.log` and show a reload recovery screen. Successful Browser Extension WebSocket upgrades also no longer appear as backend non-2xx warnings.
Graph Intelligence and AI Now recover from short model-provider connection blips before any answer has appeared. Mem retries those safe failures automatically, while rate limits and quota errors still use the normal cooldown path.
Knowledge Graph search, resize, and initial canvas render now handle asynchronous layout failures gracefully, so a graph that loaded correctly is less likely to be interrupted by a G6 layout race.
Data Transfer now includes Timeline history. Restored or cloud-imported workspaces keep past briefing, review, source-ingest, and maintenance cards instead of arriving with an empty Timeline.
AI Now keeps skill prompt injection bounded so several enabled skills or a large custom skill are less likely to overwhelm local or OpenAI-compatible models. If a selected model still stops streaming, Mem now points users to the model/provider path instead of the Nowledge service status page.

Platform

Ready to try Mem?

Download the app, connect your AI tools, and keep useful context where you control it.

Download Read Docs