Changelog

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

Latest release

0.10.83

Sep 19, 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.

0.10.83

Sep 19, 2026

0.10.83Latest releaseSep 19, 2026

Keep Work Moving

Recover conversation capture safely, follow background work clearly, and keep search and connections moving.

Highlights

Recover Codex conversation capture without losing new turns.
Browse local coding-agent sessions from Connectors.
See exactly what needs attention when search repair runs.

Added

Browse supported local coding-agent sessions directly from Connectors, then choose exactly what to import.
See the provider and model that handled each Processing task, including useful fallback and retry context.
Follow background-task progress and execution health instead of being left with an unexplained working state.

Improved

Import supported compressed Codex session snapshots.
Search repair names the affected conversations, Sources, and Memories and offers the right recovery action for each.
Watch durable reindex write progress, so a long rebuild visibly keeps moving.
Keep large Source repair results bounded, so recovery views stay responsive.
Return to the same workspace when you reopen or retry supported agent-session discovery.
See why one overlapping rule takes priority over another.
Keep preferred vocabulary during distillation review and leave syntax noise out of labels.
Search canonical Memory and Wiki documents through KFS.
Separate a skill's explicit unavailable result from relevance, so recommendations are easier to interpret.
Use the current GPT-6.5 catalog when selecting a ChatGPT Subscription model.
Restore a pending device login when you reopen the app.
Keep local projection settings available when you're offline.

Fixed

Codex capture recovers when a saved conversation already exists, without dropping a genuinely new repeated message.
ChatGPT Subscription work recovers from ordinary sign-in expiry and tells you when reconnecting is needed.
Memory extraction waits for PDF processing to finish, so a document is not distilled before its content is ready.
Same-Wi-Fi access now keeps its promised API-key protection, while ordinary localhost use remains unchanged.
Desktop startup checks only the configured backend address, so an unrelated listener no longer blocks a local Mem backend.
Working Memory rechecks your current permission and preserves data through interrupted recovery.
Manual Knowledge Graph repairs keep their admission state across interruption and cancellation.
Background task failures offer recovery instead of leaving work stuck.
Keep partial Source health visible after a retry.
Reject unsupported agent transcript formats instead of importing unreliable data.
Local imports retain parallel relationships and Space IDs.
Recover the AI Now composer after an interrupted turn.
Keep an uploaded Source visible after Library refresh.
Show OCR progress for PDF Sources.
Keep message capture independent from AI turns.

0.10.82

Sep 15, 2026

0.10.82Sep 15, 2026

Keep Every Detail

Read complete knowledge, revisit document history, and connect your AI tools to the right Mem.

Added

Adjust search wait times and retries in Settings when your embedding provider responds slowly.
Import conversations exported from Kagi Assistant as Markdown.

Improved

Browse a document's complete version history without older versions crowding the Library list.
Connector setup now uses the local or remote Mem selected in your App and guides your AI tool through connection checks.
Desktop keeps checking for updates while its window is hidden, and small graph previews pause when out of view.
Space summaries distinguish active Memories from archived items and Crystals.

Fixed

Review proposed Memory changes and their supporting evidence before confirming a Timeline decision.
Memory history corrections show the full contents being reviewed and ask you to refresh if the evidence changes.
Conversation deletion reports completed work and unfinished cleanup, keeping failed items available for follow-up.
AI Now preserves nested lists, indented code, and Markdown line breaks in model replies.
OAuth connector setup completes correctly when the provider redirects its callback address.
Background AI work respects local-only mode instead of unexpectedly contacting a remote model.
Saving an older Memory policy draft now reports a conflict instead of overwriting a newer change.
Label edits stay consistent across open lists, including recovery after a failed save.
Codex session capture preserves image attachments from current conversation records.
Background jobs stop repeating an output-budget retry when it cannot change the outcome.
Background work exposes actionable failures and bounded progress instead of remaining stuck on an unexplained working state.
Memory reviews whose pending state was left behind are reconciled so they do not remain stuck indefinitely.
AI tools receive the full text when opening a Memory, and complete conversation exports retain messages beyond the first page.
Resuming PDF OCR preserves the document's original text, continues through all remaining pages, and reuses work already completed.
Deleting a Source also removes its stored content and prevents ongoing OCR from bringing it back.
Re-uploading a coding session to an explicitly chosen Space moves the existing conversation there without creating a duplicate.
Session lifecycle capture respects your auto-sync preference, and local MCP attachment saving no longer reports a remote-only restriction.
Space selection remains consistent across the App, CLI, and AI tools, including Rules and Ontologies.
Crystal evidence loading failures show a retry action, and interrupted remote live updates are clearly identified.
Open a Memory from a standalone graph in the desktop app or its web fallback.
Re-importing an unchanged Skill bundle no longer creates a redundant update.

0.10.81

Sep 11, 2026

0.10.81Sep 11, 2026

Keep Your Work Moving

Bring in more conversation history, recover in-progress work, and keep key details readable across screens.

Highlights

Import VS Code Copilot Chat and Open WebUI conversations.
Keep unsent drafts while moving around Mem.
Read Memory details comfortably on phone-sized screens.
Find saved ChatGPT threads from their original links.

Added

Import conversation history from VS Code Copilot Chat and Open WebUI alongside the sources Mem already supports.
Choose a dedicated model for Background Intelligence, so scheduled knowledge work can use a model suited to that purpose.

Improved

Find saved ChatGPT conversations by pasting their original URL or thread ID.
Memory filters now expose the full label list instead of hiding labels beyond the initial result window.
Local coding-session discovery recognizes Devin Next databases, and Cursor imports select the real conversation file for preview and export.
The CLI preserves complete label lists across reads and shows the native embedding status in human-readable output.

Fixed

Unsent feed drafts survive navigation, so switching sections no longer discards what you were writing.
Text, controls, and sidebars stay readable in both light and dark themes.
Memory details remain readable on phone-sized screens, and compact Skill imports keep every control within reach.
Interrupted native OCR retries recover cleanly and keep the provider's error details available for diagnosis.
Restoring a backup now keeps populated source details and rebuilds relationships against items already present at the destination.
Edits to files inside an existing Skill bundle are saved reliably, while descriptions that exceed the supported size receive a clear rejection.
Conversation sync skips non-conversation artifacts and reports source failures instead of silently presenting incomplete results.
Short ChatGPT conversations sync completely even when the transcript fits inside the visible conversation area.
If ChatGPT model discovery fails, Mem keeps your saved choice and explains the error instead of offering models the account may not support.
Desktop setup passes coding-tool hook paths correctly and leaves unrelated Unix processes running during cleanup.
Windows upgrades existing Visual C++ runtimes when needed, and raw Linux packages forward sign-in and deep links into the app.

0.10.80

Sep 9, 2026

0.10.80Sep 9, 2026

Pick Up Where You Left Off

Resume supported coding sessions, see conversation sync health, and keep search and AI work on track.

Highlights

Resume a supported coding session in its original tool.
Keep imported conversations on their original dates.
Open self-hosted Web with fewer requests and smaller downloads.
Keep mobile shares safe while the connection starts.

Added

Conversation sync now shows each source's health and last successful sync, with a direct route to the settings that need attention.
OpenCode Go is now a built-in AI provider. Add your Go key in Settings or first-run setup to use its supported models in Mem.
A Memory's activity now distinguishes when an agent retrieved it, read it, or cited it in an answer, helping you see how your saved knowledge is being used.
Session previews now offer a resume command for supported coding tools, so you can reopen the same conversation in its original app. Commands appear only when Mem can identify a supported session.
The CLI and MCP tools can now save a selected Thread attachment to Library while keeping its original message link. Images remain safely saved when text recognition is unavailable.

Improved

Labs remembers your choice to hide its introductory warning, so returning to experimental settings does not repeat the same prompt.
Self-hosted Web opens with fewer startup requests and smaller downloads, and reuses downloaded app files on later visits.
Memory Maintenance does less repeated work in large libraries, and Insight Discovery moves past Memories it has already reviewed.

Fixed

When a provider rejects the model selected for Background Intelligence, Mem now explains the problem and offers a direct way to choose another model.
Stopped AI Now answers settle their unfinished tool cards, and child-task progress stops when its parent call ends instead of lingering in the conversation.
Search rebuilds retry interrupted embedding responses and no longer report completion if the text-search index could not be made ready.
Automatic Memory relationship reviews keep the saved evidence and discard outdated suggestions or contradiction flags when the underlying Memory text changes.
Thread distillation stops promptly when an AI account's quota is exhausted and keeps the provider's explanation, instead of retrying the remaining conversation chunks.
Imported Codex and Cursor conversations now keep the dates when they actually happened, so an old session no longer looks new just because it was imported today.
Current Claude account exports can now be imported from self-hosted Web and remote clients. Choose the downloaded archives you want to bring into Mem.
AI provider checks now follow the Mem you are connected to. Switching connections retires outdated setup work, remote clients avoid misleading local model-repair prompts, and an unreachable OpenAI sign-in stops waiting.
AI Now keeps custom MCP connection settings across edits and restarts, uses the connected server for authorization, and reports an unavailable tool connection before starting an answer.
Memory links stay open through startup and respect your next Space selection. Search rebuilds that stop making progress now report a failure you can act on.
Menus stay above the dialogs and panels that opened them, and ontology import controls and details remain reachable on smaller screens.
MathJax-style formulas render correctly, and missing interface translations have been restored across all nine supported languages.

Platform

Mobile shares wait for a ready connection before sending, keep failed items available for Retry, and avoid duplicate submissions during startup. Android builds now target Android 16.
Windows startup now explains when a system-reserved port prevents Mem from starting its local service, making the problem easier to diagnose.

0.10.78

Sep 5, 2026

0.10.78Sep 5, 2026

Bring the AI You Already Use

First-run setup now starts with the subscriptions and current models you already use, imported coding sessions keep their attachments, and automatic Thread review recovers after provider setup.

Highlights

Start AI setup with a subscription you already use.
Current models and regional endpoints are discovered earlier.
Automatic Thread review recovers after provider setup.
Claude Code and Codex imports keep attachment context.

Added

First-run AI setup now leads with the supported subscription routes you may already pay for. ChatGPT and GitHub Copilot sign in within the guide, while Kimi Code selects its coding-plan endpoint and asks for the matching plan key; API-key setup remains available just below.
After pairing Access Anywhere, connector setup now lets you choose Grok Bot, an eligible ChatGPT plan, or Claude directly without routing the connection through another gateway.

Improved

AI setup now checks the provider's current model list as soon as a credential is ready, prefers a current compatible model, and can find the matching global or regional endpoint automatically. You can still choose an endpoint explicitly when needed.
A fresh local Vault starts downloading the on-device search model when the setup guide opens, so the final search choice can reuse work already in progress instead of starting the download there.
Threads whose first automatic review could not run without an AI provider now get one bounded recovery attempt after a usable provider is configured.

Fixed

Claude Code and Codex session imports now retain attachment names, types, and message links for the existing explicit-save flow instead of dropping that attachment context.
Claude account exports remain recognizable when Claude changes the download URL path, so a valid export no longer falls into the ordinary JSON import path.

0.10.77

Sep 3, 2026

0.10.77Sep 3, 2026

See What Distillation Did

Complete Claude account exports now import in one guided flow, manual distillation reports its real outcome, and Library uploads keep new selections safe.

Highlights

Import complete Claude account exports, including projects and profile memory.
Manual Distill shows its real outcome from queue to completion.
A requested redistillation now does the work again.
Library uploads keep new selections safe and release large browser files.

Added

You can now import complete Claude account exports, including conversations, profile memory, and project knowledge. Choose any completed archive subset and return later for the rest; even large conversation histories import with bounded memory use.

Improved

Manual Distill now follows the task from queue to completion and reports whether it created new Memories, reinforced existing ones, found nothing new, or needs attention. It also works with an installed Local LLM and links running work to Processing.
Selecting a previously covered Thread and asking to extract it again now performs a fresh distillation. Explicit manual retries are no longer silently folded into background work.
Browser Library uploads now reuse the original file instead of making another full in-memory copy, then release it after completion, cancellation, or removal.
Library source rows now follow their actual content height, and long model lists respond to the mouse wheel and trackpad during setup.

Fixed

A file selected while another Library upload is running now stays queued for the next batch instead of disappearing when the active batch finishes.
PDF OCR now rejects malformed or mislabeled model responses instead of importing protocol or proxy text as document content.
Agent Session previews and imports now leave out hidden host scaffolding from Kimi, Codex, and OpenCode while preserving the visible conversation.
Desktop now detects local model availability correctly again, so an installed Local LLM no longer appears unavailable.
Windows startup now retries a transient executable lock instead of immediately reporting that the backend is unavailable, and Linux OAuth callbacks return to the App through the registered deep link.
Activity Calendar now marks today using your local date instead of UTC.
A search-index repair that cannot fit within its size limit now stops retrying and surfaces an actionable terminal state.
Self-hosted container installations can now activate Vault from inside the App. Authorization fails closed, and older servers show the exact nmemctl command to finish activation.

0.10.76

Sep 1, 2026

0.10.76Sep 1, 2026

Retry the Import, Stay in the Right Space

Failed imports remain recoverable, workspace views keep their scope through connection changes, and self-hosted MCP sign-in returns to the browser that started it.

Highlights

Failed conversation imports remain retryable.
Feed, Tree, and Files stay in the selected Space.
Self-hosted MCP OAuth returns to the browser that started it.
Personal data-location choices now complete safely.

Improved

Freshly imported conversations become searchable from their source Threads before Memory extraction finishes. An answer still inspects the matching transcript before relying on it.
Thread readers now preserve original message timestamps in nmem CLI and MCP results, making imported session timelines easier to reconstruct accurately.

Fixed

A failed conversation import stays selected and retryable instead of looking finished. During first-time Team setup, the setup window also stays open while the workspace connection completes.
Feed, Tree, and Files now keep reads tied to the Space you selected. Rapid Space or backend switches discard stale results instead of showing data or local diagnostics from the previous view.
Self-hosted Web MCP connections now complete OAuth on the same browser origin, while copied hosted MCP configuration includes the active credential and fails closed if the connection changes mid-copy.
The Personal data-location choice announced in 0.10.75 now ships in the App and follows through safely: keep data on this device or open Personal Cloud without falling into a Team workspace or stranding the device copy.
OpenRouter vision-capable models now appear in the Library selector, and AI Now clears a finished background session from the running indicator even after you switch conversations.
The Library Crystal tab now shows every retained Crystal, including archived or historical knowledge pages that remain valid but are no longer the latest Memory version.
Conversation Markdown downloaded by the browser extension now imports as the original User and Assistant turns instead of collapsing the whole file into one document message.
Opening a Feed Memory in Graph now keeps its originating Space, so the focused Memory and its neighbors no longer fall back to another workspace view.
Agent Sessions now keeps project selection controls aligned with project names and contains the checkmark inside its checkbox, including rows that show a workspace path.
Default mode now opens the sidebar with advanced destinations still folded under More, keeping the focused Home navigation quiet until you choose to expand it.

0.10.75

Aug 31, 2026

0.10.75Aug 31, 2026

A First Answer You Can Trust

First-run setup now waits for the work you selected and a working AI connection, Agent Sessions make each workspace easy to identify, and remote health checks reveal less without credentials.

Highlights

The first answer waits for real import and AI readiness.
Full workspace paths make Agent Sessions easier to navigate.
Team setup keeps your first imported conversations in the connected workspace.
Remote health checks share less information before sign-in.

Improved

Your first question after setup now waits until selected conversation imports have actually finished and AI is ready. If an import fails, you can retry it; when no history was imported, Mem says so and answers from a read-only guide instead of inventing details.
Agent Sessions now shows the full workspace path beneath each project, so similarly named worktrees are easy to tell apart. Long paths stay quiet in the list while remaining available to copy.
Team members who sign in during first-time setup stay connected to their Team workspace as setup continues, so selected conversations are imported into the place they expect.

Fixed

Remote health checks now return only a simple liveness result unless you are authenticated, keeping storage, model, and graph diagnostics out of unauthenticated responses. Trusted local status views and nmem doctor still show the details needed to troubleshoot your installation.

0.10.74

Aug 29, 2026

0.10.74Aug 29, 2026

Set Up AI in Place, and Keep What Already Worked

First-run setup can finish your AI connection without sending you elsewhere, AI Now is clearer about what it retrieved, and partial failures preserve the useful work that already succeeded.

Highlights

Choose and verify an AI provider without leaving first-run setup.
AI Now searches broadly before opening only the strongest matches.
Readable PDF pages survive a damaged page or fallback-parser failure.
Background tabs pause interface polling without stopping backend work.

Added

You can now choose an AI provider, verify its credential, save the model, and start using it without leaving first-run setup. ChatGPT sign-in remains available as a separate option.
Links to a specific Memory or Thread now open that item directly in the web app, while keeping the same access rules as the rest of your workspace.

Improved

AI Now checks the relevant Memory, Thread, and Library collections together, then opens only the strongest matches for detail. Search hits and inspected details remain visibly distinct.
When Mem is hidden in the background, the web interface pauses its own status checks and display timers, then resumes normally when you return. Ongoing backend work continues uninterrupted.
Skill suggestions and compiled Skills now keep using the output language selected in your Profile, even when the supporting material is written in another language.

Fixed

Local Library PDF imports now keep the readable pages when another page is damaged or a fallback parser cannot open the document. Failed pages stay in their correct positions for later repair.
If a local knowledge-graph extraction batch partly fails, Retry now reruns only the failed Memories instead of repeating the entire batch.
AI providers that send a tool name late in a Responses stream can now complete that tool call once, instead of losing it or running it twice.
On mobile, an insecure remote HTTP address now explains how to switch to trusted HTTPS or Tailscale Serve. Local loopback development remains available.

0.10.73

Aug 27, 2026

0.10.73Aug 27, 2026

Tools Connect Faster, and Agent Threads Stay Readable

Setup can now connect supported AI tools and verify the result, agent transcripts keep tool work tucked away until you need it, and capture and account edge cases recover more honestly.

Highlights

Connect Claude Code or Codex from Setup and verify the real configuration in one step.
Tool work stays collapsed in agent transcripts until you need its details.
DeepSeek capture now stores the final answer instead of its reasoning block.
Accounts without a hosted workspace keep capturing Memories locally.

Added

Setup now detects installed AI tools and offers one-click Connect for Claude Code and Codex. It only shows success after checking the real configuration; other supported tools keep their guided manual path.
Thread detail now groups each tool call and result into a compact, collapsed unit. Expand it when you need the input, output, or error without letting execution details take over the conversation.

Improved

Local-only model providers remain available for conversation but are now clearly marked Local Only and unavailable for Memory extraction.
When two Skills resolve to the same name or path, Mem now reports the collision instead of silently adding an ambiguous managed copy.
Memory Health now says when a completed pass found no type changes, rather than making a successful no-op sound unfinished.

Fixed

DeepSeek browser capture now keeps the final answer and skips the model's reasoning block.
Signing in without a hosted workspace no longer breaks local Memory capture. Writes stay local until there is a remote workspace to receive them.
AI Now now replaces refreshed OAuth credentials in the active session, uses a next-task Identity for exactly one successful task, and clears a selected Identity after it is authoritatively deleted.
On Windows, updater relaunch and enabled autostart now follow the packaged executable's real path while preserving an autostart choice that Windows has disabled.
If a remote Thread remains missing, recovery now stops after three failed attempts and shows the retry state instead of looping forever.

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

0.10.83

Sep 19, 2026

0.10 is a ground-up engine rewrite

0.10.83

Sep 19, 2026

0.10.83Latest releaseSep 19, 2026

Keep Work Moving

Recover conversation capture safely, follow background work clearly, and keep search and connections moving.

Highlights

Recover Codex conversation capture without losing new turns.
Browse local coding-agent sessions from Connectors.
See exactly what needs attention when search repair runs.

Added

Browse supported local coding-agent sessions directly from Connectors, then choose exactly what to import.
See the provider and model that handled each Processing task, including useful fallback and retry context.
Follow background-task progress and execution health instead of being left with an unexplained working state.

Improved

Import supported compressed Codex session snapshots.
Search repair names the affected conversations, Sources, and Memories and offers the right recovery action for each.
Watch durable reindex write progress, so a long rebuild visibly keeps moving.
Keep large Source repair results bounded, so recovery views stay responsive.
Return to the same workspace when you reopen or retry supported agent-session discovery.
See why one overlapping rule takes priority over another.
Keep preferred vocabulary during distillation review and leave syntax noise out of labels.
Search canonical Memory and Wiki documents through KFS.
Separate a skill's explicit unavailable result from relevance, so recommendations are easier to interpret.
Use the current GPT-6.5 catalog when selecting a ChatGPT Subscription model.
Restore a pending device login when you reopen the app.
Keep local projection settings available when you're offline.

Fixed

Codex capture recovers when a saved conversation already exists, without dropping a genuinely new repeated message.
ChatGPT Subscription work recovers from ordinary sign-in expiry and tells you when reconnecting is needed.
Memory extraction waits for PDF processing to finish, so a document is not distilled before its content is ready.
Same-Wi-Fi access now keeps its promised API-key protection, while ordinary localhost use remains unchanged.
Desktop startup checks only the configured backend address, so an unrelated listener no longer blocks a local Mem backend.
Working Memory rechecks your current permission and preserves data through interrupted recovery.
Manual Knowledge Graph repairs keep their admission state across interruption and cancellation.
Background task failures offer recovery instead of leaving work stuck.
Keep partial Source health visible after a retry.
Reject unsupported agent transcript formats instead of importing unreliable data.
Local imports retain parallel relationships and Space IDs.
Recover the AI Now composer after an interrupted turn.
Keep an uploaded Source visible after Library refresh.
Show OCR progress for PDF Sources.
Keep message capture independent from AI turns.

0.10.82

Sep 15, 2026

0.10.82Sep 15, 2026

Keep Every Detail

Read complete knowledge, revisit document history, and connect your AI tools to the right Mem.

Added

Adjust search wait times and retries in Settings when your embedding provider responds slowly.
Import conversations exported from Kagi Assistant as Markdown.

Improved

Browse a document's complete version history without older versions crowding the Library list.
Connector setup now uses the local or remote Mem selected in your App and guides your AI tool through connection checks.
Desktop keeps checking for updates while its window is hidden, and small graph previews pause when out of view.
Space summaries distinguish active Memories from archived items and Crystals.

Fixed

Review proposed Memory changes and their supporting evidence before confirming a Timeline decision.
Memory history corrections show the full contents being reviewed and ask you to refresh if the evidence changes.
Conversation deletion reports completed work and unfinished cleanup, keeping failed items available for follow-up.
AI Now preserves nested lists, indented code, and Markdown line breaks in model replies.
OAuth connector setup completes correctly when the provider redirects its callback address.
Background AI work respects local-only mode instead of unexpectedly contacting a remote model.
Saving an older Memory policy draft now reports a conflict instead of overwriting a newer change.
Label edits stay consistent across open lists, including recovery after a failed save.
Codex session capture preserves image attachments from current conversation records.
Background jobs stop repeating an output-budget retry when it cannot change the outcome.
Background work exposes actionable failures and bounded progress instead of remaining stuck on an unexplained working state.
Memory reviews whose pending state was left behind are reconciled so they do not remain stuck indefinitely.
AI tools receive the full text when opening a Memory, and complete conversation exports retain messages beyond the first page.
Resuming PDF OCR preserves the document's original text, continues through all remaining pages, and reuses work already completed.
Deleting a Source also removes its stored content and prevents ongoing OCR from bringing it back.
Re-uploading a coding session to an explicitly chosen Space moves the existing conversation there without creating a duplicate.
Session lifecycle capture respects your auto-sync preference, and local MCP attachment saving no longer reports a remote-only restriction.
Space selection remains consistent across the App, CLI, and AI tools, including Rules and Ontologies.
Crystal evidence loading failures show a retry action, and interrupted remote live updates are clearly identified.
Open a Memory from a standalone graph in the desktop app or its web fallback.
Re-importing an unchanged Skill bundle no longer creates a redundant update.

0.10.81

Sep 11, 2026

0.10.81Sep 11, 2026

Keep Your Work Moving

Bring in more conversation history, recover in-progress work, and keep key details readable across screens.

Highlights

Import VS Code Copilot Chat and Open WebUI conversations.
Keep unsent drafts while moving around Mem.
Read Memory details comfortably on phone-sized screens.
Find saved ChatGPT threads from their original links.

Added

Import conversation history from VS Code Copilot Chat and Open WebUI alongside the sources Mem already supports.
Choose a dedicated model for Background Intelligence, so scheduled knowledge work can use a model suited to that purpose.

Improved

Find saved ChatGPT conversations by pasting their original URL or thread ID.
Memory filters now expose the full label list instead of hiding labels beyond the initial result window.
Local coding-session discovery recognizes Devin Next databases, and Cursor imports select the real conversation file for preview and export.
The CLI preserves complete label lists across reads and shows the native embedding status in human-readable output.

Fixed

Unsent feed drafts survive navigation, so switching sections no longer discards what you were writing.
Text, controls, and sidebars stay readable in both light and dark themes.
Memory details remain readable on phone-sized screens, and compact Skill imports keep every control within reach.
Interrupted native OCR retries recover cleanly and keep the provider's error details available for diagnosis.
Restoring a backup now keeps populated source details and rebuilds relationships against items already present at the destination.
Edits to files inside an existing Skill bundle are saved reliably, while descriptions that exceed the supported size receive a clear rejection.
Conversation sync skips non-conversation artifacts and reports source failures instead of silently presenting incomplete results.
Short ChatGPT conversations sync completely even when the transcript fits inside the visible conversation area.
If ChatGPT model discovery fails, Mem keeps your saved choice and explains the error instead of offering models the account may not support.
Desktop setup passes coding-tool hook paths correctly and leaves unrelated Unix processes running during cleanup.
Windows upgrades existing Visual C++ runtimes when needed, and raw Linux packages forward sign-in and deep links into the app.

0.10.80

Sep 9, 2026

0.10.80Sep 9, 2026

Pick Up Where You Left Off

Resume supported coding sessions, see conversation sync health, and keep search and AI work on track.

Highlights

Resume a supported coding session in its original tool.
Keep imported conversations on their original dates.
Open self-hosted Web with fewer requests and smaller downloads.
Keep mobile shares safe while the connection starts.

Added

Conversation sync now shows each source's health and last successful sync, with a direct route to the settings that need attention.
OpenCode Go is now a built-in AI provider. Add your Go key in Settings or first-run setup to use its supported models in Mem.
A Memory's activity now distinguishes when an agent retrieved it, read it, or cited it in an answer, helping you see how your saved knowledge is being used.
Session previews now offer a resume command for supported coding tools, so you can reopen the same conversation in its original app. Commands appear only when Mem can identify a supported session.
The CLI and MCP tools can now save a selected Thread attachment to Library while keeping its original message link. Images remain safely saved when text recognition is unavailable.

Improved

Labs remembers your choice to hide its introductory warning, so returning to experimental settings does not repeat the same prompt.
Self-hosted Web opens with fewer startup requests and smaller downloads, and reuses downloaded app files on later visits.
Memory Maintenance does less repeated work in large libraries, and Insight Discovery moves past Memories it has already reviewed.

Fixed

When a provider rejects the model selected for Background Intelligence, Mem now explains the problem and offers a direct way to choose another model.
Stopped AI Now answers settle their unfinished tool cards, and child-task progress stops when its parent call ends instead of lingering in the conversation.
Search rebuilds retry interrupted embedding responses and no longer report completion if the text-search index could not be made ready.
Automatic Memory relationship reviews keep the saved evidence and discard outdated suggestions or contradiction flags when the underlying Memory text changes.
Thread distillation stops promptly when an AI account's quota is exhausted and keeps the provider's explanation, instead of retrying the remaining conversation chunks.
Imported Codex and Cursor conversations now keep the dates when they actually happened, so an old session no longer looks new just because it was imported today.
Current Claude account exports can now be imported from self-hosted Web and remote clients. Choose the downloaded archives you want to bring into Mem.
AI provider checks now follow the Mem you are connected to. Switching connections retires outdated setup work, remote clients avoid misleading local model-repair prompts, and an unreachable OpenAI sign-in stops waiting.
AI Now keeps custom MCP connection settings across edits and restarts, uses the connected server for authorization, and reports an unavailable tool connection before starting an answer.
Memory links stay open through startup and respect your next Space selection. Search rebuilds that stop making progress now report a failure you can act on.
Menus stay above the dialogs and panels that opened them, and ontology import controls and details remain reachable on smaller screens.
MathJax-style formulas render correctly, and missing interface translations have been restored across all nine supported languages.

Platform

Mobile shares wait for a ready connection before sending, keep failed items available for Retry, and avoid duplicate submissions during startup. Android builds now target Android 16.
Windows startup now explains when a system-reserved port prevents Mem from starting its local service, making the problem easier to diagnose.

0.10.78

Sep 5, 2026

0.10.78Sep 5, 2026

Bring the AI You Already Use

First-run setup now starts with the subscriptions and current models you already use, imported coding sessions keep their attachments, and automatic Thread review recovers after provider setup.

Highlights

Start AI setup with a subscription you already use.
Current models and regional endpoints are discovered earlier.
Automatic Thread review recovers after provider setup.
Claude Code and Codex imports keep attachment context.

Added

First-run AI setup now leads with the supported subscription routes you may already pay for. ChatGPT and GitHub Copilot sign in within the guide, while Kimi Code selects its coding-plan endpoint and asks for the matching plan key; API-key setup remains available just below.
After pairing Access Anywhere, connector setup now lets you choose Grok Bot, an eligible ChatGPT plan, or Claude directly without routing the connection through another gateway.

Improved

AI setup now checks the provider's current model list as soon as a credential is ready, prefers a current compatible model, and can find the matching global or regional endpoint automatically. You can still choose an endpoint explicitly when needed.
A fresh local Vault starts downloading the on-device search model when the setup guide opens, so the final search choice can reuse work already in progress instead of starting the download there.
Threads whose first automatic review could not run without an AI provider now get one bounded recovery attempt after a usable provider is configured.

Fixed

Claude Code and Codex session imports now retain attachment names, types, and message links for the existing explicit-save flow instead of dropping that attachment context.
Claude account exports remain recognizable when Claude changes the download URL path, so a valid export no longer falls into the ordinary JSON import path.

0.10.77

Sep 3, 2026

0.10.77Sep 3, 2026

See What Distillation Did

Complete Claude account exports now import in one guided flow, manual distillation reports its real outcome, and Library uploads keep new selections safe.

Highlights

Import complete Claude account exports, including projects and profile memory.
Manual Distill shows its real outcome from queue to completion.
A requested redistillation now does the work again.
Library uploads keep new selections safe and release large browser files.

Added

You can now import complete Claude account exports, including conversations, profile memory, and project knowledge. Choose any completed archive subset and return later for the rest; even large conversation histories import with bounded memory use.

Improved

Manual Distill now follows the task from queue to completion and reports whether it created new Memories, reinforced existing ones, found nothing new, or needs attention. It also works with an installed Local LLM and links running work to Processing.
Selecting a previously covered Thread and asking to extract it again now performs a fresh distillation. Explicit manual retries are no longer silently folded into background work.
Browser Library uploads now reuse the original file instead of making another full in-memory copy, then release it after completion, cancellation, or removal.
Library source rows now follow their actual content height, and long model lists respond to the mouse wheel and trackpad during setup.

Fixed

A file selected while another Library upload is running now stays queued for the next batch instead of disappearing when the active batch finishes.
PDF OCR now rejects malformed or mislabeled model responses instead of importing protocol or proxy text as document content.
Agent Session previews and imports now leave out hidden host scaffolding from Kimi, Codex, and OpenCode while preserving the visible conversation.
Desktop now detects local model availability correctly again, so an installed Local LLM no longer appears unavailable.
Windows startup now retries a transient executable lock instead of immediately reporting that the backend is unavailable, and Linux OAuth callbacks return to the App through the registered deep link.
Activity Calendar now marks today using your local date instead of UTC.
A search-index repair that cannot fit within its size limit now stops retrying and surfaces an actionable terminal state.
Self-hosted container installations can now activate Vault from inside the App. Authorization fails closed, and older servers show the exact nmemctl command to finish activation.

0.10.76

Sep 1, 2026

0.10.76Sep 1, 2026

Retry the Import, Stay in the Right Space

Failed imports remain recoverable, workspace views keep their scope through connection changes, and self-hosted MCP sign-in returns to the browser that started it.

Highlights

Failed conversation imports remain retryable.
Feed, Tree, and Files stay in the selected Space.
Self-hosted MCP OAuth returns to the browser that started it.
Personal data-location choices now complete safely.

Improved

Freshly imported conversations become searchable from their source Threads before Memory extraction finishes. An answer still inspects the matching transcript before relying on it.
Thread readers now preserve original message timestamps in nmem CLI and MCP results, making imported session timelines easier to reconstruct accurately.

Fixed

A failed conversation import stays selected and retryable instead of looking finished. During first-time Team setup, the setup window also stays open while the workspace connection completes.
Feed, Tree, and Files now keep reads tied to the Space you selected. Rapid Space or backend switches discard stale results instead of showing data or local diagnostics from the previous view.
Self-hosted Web MCP connections now complete OAuth on the same browser origin, while copied hosted MCP configuration includes the active credential and fails closed if the connection changes mid-copy.
The Personal data-location choice announced in 0.10.75 now ships in the App and follows through safely: keep data on this device or open Personal Cloud without falling into a Team workspace or stranding the device copy.
OpenRouter vision-capable models now appear in the Library selector, and AI Now clears a finished background session from the running indicator even after you switch conversations.
The Library Crystal tab now shows every retained Crystal, including archived or historical knowledge pages that remain valid but are no longer the latest Memory version.
Conversation Markdown downloaded by the browser extension now imports as the original User and Assistant turns instead of collapsing the whole file into one document message.
Opening a Feed Memory in Graph now keeps its originating Space, so the focused Memory and its neighbors no longer fall back to another workspace view.
Agent Sessions now keeps project selection controls aligned with project names and contains the checkmark inside its checkbox, including rows that show a workspace path.
Default mode now opens the sidebar with advanced destinations still folded under More, keeping the focused Home navigation quiet until you choose to expand it.

0.10.75

Aug 31, 2026

0.10.75Aug 31, 2026

A First Answer You Can Trust

First-run setup now waits for the work you selected and a working AI connection, Agent Sessions make each workspace easy to identify, and remote health checks reveal less without credentials.

Highlights

The first answer waits for real import and AI readiness.
Full workspace paths make Agent Sessions easier to navigate.
Team setup keeps your first imported conversations in the connected workspace.
Remote health checks share less information before sign-in.

Improved

Your first question after setup now waits until selected conversation imports have actually finished and AI is ready. If an import fails, you can retry it; when no history was imported, Mem says so and answers from a read-only guide instead of inventing details.
Agent Sessions now shows the full workspace path beneath each project, so similarly named worktrees are easy to tell apart. Long paths stay quiet in the list while remaining available to copy.
Team members who sign in during first-time setup stay connected to their Team workspace as setup continues, so selected conversations are imported into the place they expect.

Fixed

Remote health checks now return only a simple liveness result unless you are authenticated, keeping storage, model, and graph diagnostics out of unauthenticated responses. Trusted local status views and nmem doctor still show the details needed to troubleshoot your installation.

0.10.74

Aug 29, 2026

0.10.74Aug 29, 2026

Set Up AI in Place, and Keep What Already Worked

First-run setup can finish your AI connection without sending you elsewhere, AI Now is clearer about what it retrieved, and partial failures preserve the useful work that already succeeded.

Highlights

Choose and verify an AI provider without leaving first-run setup.
AI Now searches broadly before opening only the strongest matches.
Readable PDF pages survive a damaged page or fallback-parser failure.
Background tabs pause interface polling without stopping backend work.

Added

You can now choose an AI provider, verify its credential, save the model, and start using it without leaving first-run setup. ChatGPT sign-in remains available as a separate option.
Links to a specific Memory or Thread now open that item directly in the web app, while keeping the same access rules as the rest of your workspace.

Improved

AI Now checks the relevant Memory, Thread, and Library collections together, then opens only the strongest matches for detail. Search hits and inspected details remain visibly distinct.
When Mem is hidden in the background, the web interface pauses its own status checks and display timers, then resumes normally when you return. Ongoing backend work continues uninterrupted.
Skill suggestions and compiled Skills now keep using the output language selected in your Profile, even when the supporting material is written in another language.

Fixed

Local Library PDF imports now keep the readable pages when another page is damaged or a fallback parser cannot open the document. Failed pages stay in their correct positions for later repair.
If a local knowledge-graph extraction batch partly fails, Retry now reruns only the failed Memories instead of repeating the entire batch.
AI providers that send a tool name late in a Responses stream can now complete that tool call once, instead of losing it or running it twice.
On mobile, an insecure remote HTTP address now explains how to switch to trusted HTTPS or Tailscale Serve. Local loopback development remains available.

0.10.73

Aug 27, 2026

0.10.73Aug 27, 2026

Tools Connect Faster, and Agent Threads Stay Readable

Setup can now connect supported AI tools and verify the result, agent transcripts keep tool work tucked away until you need it, and capture and account edge cases recover more honestly.

Highlights

Connect Claude Code or Codex from Setup and verify the real configuration in one step.
Tool work stays collapsed in agent transcripts until you need its details.
DeepSeek capture now stores the final answer instead of its reasoning block.
Accounts without a hosted workspace keep capturing Memories locally.

Added

Setup now detects installed AI tools and offers one-click Connect for Claude Code and Codex. It only shows success after checking the real configuration; other supported tools keep their guided manual path.
Thread detail now groups each tool call and result into a compact, collapsed unit. Expand it when you need the input, output, or error without letting execution details take over the conversation.

Improved

Local-only model providers remain available for conversation but are now clearly marked Local Only and unavailable for Memory extraction.
When two Skills resolve to the same name or path, Mem now reports the collision instead of silently adding an ambiguous managed copy.
Memory Health now says when a completed pass found no type changes, rather than making a successful no-op sound unfinished.

Fixed

DeepSeek browser capture now keeps the final answer and skips the model's reasoning block.
Signing in without a hosted workspace no longer breaks local Memory capture. Writes stay local until there is a remote workspace to receive them.
AI Now now replaces refreshed OAuth credentials in the active session, uses a next-task Identity for exactly one successful task, and clears a selected Identity after it is authoritatively deleted.
On Windows, updater relaunch and enabled autostart now follow the packaged executable's real path while preserving an autostart choice that Windows has disabled.
If a remote Thread remains missing, recovery now stops after three failed attempts and shows the retry state instead of looping forever.

Ready to try Mem?

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

Download Read Docs