Skip to main content

Changelog

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

Latest release
0.10.113
Nov 3, 2026
0.10.113Latest releaseNov 3, 2026

Clearer Skills and More Reliable Conversation Recovery

Continue Skill improvements in AI Now, inspect recorded AI usage, and import recovered conversations with fewer interruptions.

Highlights
  • Skills with AI Now editing available have a simpler detail view, a single improvement action and quick access to recent improvement conversations.
  • Inspect every AI usage source recorded by this Mem instance by time period and activity. Calls without reported token counts remain visible.
  • Previously recovered Codex archives, including compressed archives, can be imported again without duplicate messages when their complete source history is verified.
Improved
  • Skills with AI Now editing available have a simpler detail view, a single improvement action and quick access to recent improvement conversations.
  • Inspect every AI usage source recorded by this Mem instance by time period and activity. Calls without reported token counts remain visible.
  • The built-in local model uses a working window sized to your computer's memory for conversation distillation. CPU-only setups explain the expected slowness and available model alternatives.
  • Conversation and file imports avoid additional temporary message copies while preserving imported content, order and retry behavior.
  • Access Anywhere clearly warns that Quick link addresses change after every restart and points to persistent connection options.
Fixed
  • AI Now refreshes a saved Skill proposal so Apply and Dismiss appear immediately. The sidebar editor uses the available height and opens full details while retaining your draft.
  • Recovering AI Now history carries the visible conversation and unsent request into an editable draft without automatically resending it.
  • Previously recovered Codex archives, including compressed archives, can be imported again without duplicate messages when their complete source history is verified.
  • Reading messages with unusually long Markdown runs stays more responsive while preserving formatting and complete copy and export content.
  • Conversation distillation skips legacy developer instructions while preserving the original saved transcript.
  • Distillation can recover replies that run out of output space on slower providers with a fresh request window; recovery sizing uses reported output evidence.
Platform
  • The embedded database is updated to HawDB 0.25.0. Runtime admission remains off by default for datasets that have never enabled it.
  • The backend's live API documentation reports the version of the running backend.
0.10.112Nov 1, 2026

Safer Conversation Actions and Lighter Imports

Keep conversation actions tied to the selected saved history, import large archives with less memory, and preserve model settings through edits.

Highlights
  • Conversation reads, exports and handoffs retain the selected saved conversation and Space when public IDs overlap. Ambiguous references offer recovery instead of choosing another row.
  • Proma and large local Agent Session imports use less peak memory while preserving complete messages, selection order and failed-session retry.
  • Access Anywhere no longer mistakenly blocks browser-based MCP clients' public OAuth discovery, registration and token requests. Protected routes retain their origin checks.
Improved
  • Proma and large local Agent Session imports use less peak memory while preserving complete messages, selection order and failed-session retry.
  • Large ZIP and direct-file imports stop at their read limits and report oversized content. Healthy parsed content remains available where supported, and processed payloads are released sooner.
  • Claude account imports can continue with healthy conversations and project documents when profile memories are oversized. Import previews distinguish available content from completed imports.
  • Distillation failure details identify the rule and draft or judgment that failed validation, without recording the model reply text.
  • With Identity Graph enabled, removing a profile hides it from active lists while keeping saved Memories and confirmed references intact.
  • Antigravity subagent conversations sync as separate Threads when their recorded delegation can be read. Delegated tasks remain system messages, and unavailable lineage is explained without an unhelpful retry prompt.
  • Access Anywhere provides a connector file with the sign-in and token addresses that hosted agents need, reducing setup errors caused by manually retyping long URLs.
Fixed
  • Access Anywhere no longer mistakenly blocks browser-based MCP clients' public OAuth discovery, registration and token requests. Protected routes retain their origin checks.
  • Conversation reads, exports and handoffs retain the selected saved conversation and Space when public IDs overlap. Ambiguous references offer recovery instead of choosing another row.
  • Conversation actions recheck the saved identity before changes. Retrying a partial bulk deletion keeps excluded conversations out of the retry.
  • Moving a conversation retires pending actions for its previous Space. Reopening active distillation retains its confirmation, task and result tracking.
  • Stopped search repairs stay tied to the correct saved conversation, with read-only message inspection in its original Space.
  • Assigning Magpie models to Fast decisions preserves saved provider settings and unsaved edits. Reasoning choices and an independent Default model survive unrelated saves; conflicting writes share pending controls.
  • Stats and onboarding retain their last successful values when background refresh fails, and recover without a disruptive retry prompt.
  • Desktop and self-hosted local Spaces show Delete and Merge actions again. Contextual Edit and Delete labels are shorter and clearer.
  • Threads keeps keyboard focus on the chosen control after navigation.
  • Process diagnostics show CPU measurements as unavailable when a translated Mac process has unverified clock units, while retaining process identity and memory readings.
  • macOS updates started from this version tolerate unrelated SMB mount warnings during the backend port check, while real process-inspection failures still stop the update.
  • Full-text index rebuilds check available disk space before replacing the final index. If space is insufficient, the existing index is kept and the rebuild can be retried after freeing space.
  • Switching AI modes and restoring settings preserve working model connections and saved selections. Incomplete routes remain visible with repair guidance and skipped-connection warnings.
  • Connection-status refreshes no longer overwrite the saved OAuth authorization state.
  • Memory Health shows the recorded Keep for now result when the server confirms that Memories were kept, even if its optional cooldown flag is absent.
  • Timeline evidence reviews close when a refresh confirms that all selected decisions were already recorded, without submitting those decisions again.
  • Local Working Memory and AI startup context served by the app no longer treat old memory-and-thread count summaries as the current briefing. Genuine briefings and additional content remain readable, including in past Timeline entries.
  • Codex shell commands identified by the host no longer become conversation content or Thread titles. Ordinary messages quoting shell-command output remain readable.
  • Skill evidence updates preserve timestamps when counting the days on which evidence occurred.
  • Invalid AI provider settings requests return consistent JSON error details.
0.10.111Oct 30, 2026

Reliable Remote Access, Selection, and Graph Editing

Connect remotely, select conversations across pages, and keep graph drafts and context previews tied to your current work.

Highlights
  • Access Anywhere accepts HTTPS clients and remote authorization pages that could previously be rejected.
  • Select all conversations matching a source filter across pages. Incomplete reads show an error; changing the selection or view retires pending requests.
  • Graph relation drafts retain their name and reason after a failed save. An earlier save finishing later no longer clears a reopened draft.
Improved
  • Graph controls work with the keyboard and remain reachable in narrow views. Hand mode keeps panning, path failures offer retry, and clearing or replacing a path selection prevents old results from taking over.
  • AI Now file downloads are visible to keyboard and touch users. File and Skill previews remain readable beside the sidebar.
  • Expandable details have clearer controls and keyboard focus. Label cleanup keeps retry and recovery actions reachable in smaller windows.
  • Choose 10, 30, 50 or 100 conversations per Threads page. The choice is remembered on this device.
  • When Codex sessions are stored in an isolated runtime home, the CLI explains how to select their directory with --session-dir and clarifies what --all-projects covers.
  • Custom Chat Completions providers retain Auto reasoning after saving, keep other generation settings when edited, and support --reasoning-effort in headless configuration.
  • Proma discovery reads one transcript record at a time and retains readable sessions when another file fails. Failed files offer retry; batch CLI imports report partial failure after processing healthy sessions.
  • Memory Health explains source-check warnings and the effect of each keep action before you choose, with clearer compact cards and mobile controls.
  • AI Identity naming is consistent across Team, Context and Settings, while tool access and model configuration stay the same.
  • Headless decision-budget errors explain which settings need attention, the cost units, and the next setup steps.
Fixed
  • Access Anywhere accepts HTTPS clients and remote authorization pages that could previously be rejected.
  • Select all conversations matching a source filter across pages. Incomplete reads show an error; changing the selection or view retires pending requests.
  • Graph relation drafts retain their name and reason after a failed save. An earlier save finishing later no longer clears a reopened draft.
  • Context Preview and copied AI context follow the current identity or connection. Delayed requests no longer substitute earlier instructions, and failed reads remain retryable.
  • Crystal list previews stay compact in browsers without line-clamp support. Opening a Crystal still shows its full content.
  • When first-run history import fails, see a reason and next step for each session. Retrying imports only failed sessions and preserves successful imports.
  • Thread details indicate when some related Memories could not be included, while keeping available Memories visible.
  • Startup identifies database shadow files whose main database is missing and explains how to prepare a separate recovery copy. Original files remain in place.
  • Recovery notices on Web and remote connections show safe location summaries while retaining recovery and retry guidance. Full host paths remain available in local desktop diagnostics.
  • Thread messages and AI Now preserve literal Nowledge resource URIs inside code and existing links, while prose keeps compact resource links.
  • Wiki keyboard actions open the intended Crystal, Entity or topic. Tree navigation follows the focused row, and selecting summary text keeps the current detail open.
  • Access Anywhere connector approval now completes after Allow instead of returning an access-denied error. Approvals remain tied to the permissions reviewed and retryable after temporary save failures.
  • Mem refuses a write when its checkpoint retry newly exhausts graph memory, preserving the evidence needed for recovery.
  • Move conversations between Spaces on non-Team Desktop and self-hosted servers again. Destination names stay consistent across menus, confirmations and success messages.
  • When a Codex archive cannot replace referenced saved history, the review explains that history is unchanged and offers a clear dismissal.
  • Codex image requests remain readable when imported attachment details start after a blank line. Copying a message preserves its original imported text.
  • Session-discovery errors use translated messages. Web and remote views hide host filenames and raw diagnostics; local Desktop can identify the affected file.
Platform
  • Restore rotating busy indicators on macOS and Linux, and refine the Mac installer artwork and layout.
0.10.110Oct 28, 2026

Smoother Updates, Reliable Capture, and Clearer Skill Control

Update in place, keep conversation capture recoverable, and review Skills and previews with clearer controls.

Highlights
  • Windows installers update existing installations in place and ask a running App to quit before making changes. Install and uninstall cleanup follows the installed-file manifest, removing obsolete App files while preserving files the App did not install.
  • Codex history recovery preserves source-backed messages without duplicates. Manual saves and bulk sync can continue with automatic capture; older imports may need one explicit recovery. Legacy image details are added only when they match a saved message unambiguously; image files are not copied.
  • Compare duplicate Skills side by side, retry failed details and see what Keep both preserves. Normal instructions after horizontal rules remain visible, and creation fields and guidance are easier to identify.
Added
  • Choose a Space when starting an AI Now conversation; existing conversations keep their original Space. AI Now can enable, turn off and remove Skills with explicit permission choices, including a chat-scoped choice for repeated enabling. The conversation list follows the selected Space; All Spaces shows the full list.
  • Import ordinary Cherry Studio 2.1.8 chats alongside Agent sessions. Selected alternative branches stay in separate Threads, while trashed and internal Agent sessions are excluded.
Improved
  • Keep plugin capture enabled while local session-file scanning is off. The scan preference survives restarting the App, with recovery actions when a setting or stop request cannot be confirmed.
  • Compare duplicate Skills side by side, retry failed details and see what Keep both preserves. Normal instructions after horizontal rules remain visible, and creation fields and guidance are easier to identify.
  • Browse saved conversations by project with complete counts and distinct folder paths. Review backup contents and warnings before restoring, with protection against outdated selections.
  • Use the keyboard with Thread cards, floating Memory search, companion launchers and Ontology diagrams. Cancelled drags stop cleanly, while rename and search inputs preserve focus and input-method composition.
  • Memory search and lists read less data before showing results, with fewer repeated graph lookups.
  • AI Now process diagnostics distinguish the CPU sample duration, per-core usage and total CPU time.
Fixed
  • Codex history recovery preserves source-backed messages without duplicates. Manual saves and bulk sync can continue with automatic capture; older imports may need one explicit recovery. Legacy image details are added only when they match a saved message unambiguously; image files are not copied.
  • Memory previews can close while loading and retry failed reads, including offline local-backend access. Copy success follows the clipboard write. Ontology previews keep the current draft, and schedule drafts survive language changes and saving.
  • Reusing a renamed label preserves existing assignments, and invalid replacements keep the prior label set. Thread refreshes withdraw full messages when access is reduced while preserving the loaded window when access remains available.
  • Restarted ChatGPT sign-ins ignore older callbacks, and cleared MCP activity stays cleared when an old history request completes. Native model catalogs preserve exact provider model IDs, with accurate AI Now labels in CLI confirmations.
  • Background work has an appropriate minimum request deadline and avoids retries that cannot finish in time. Failure details distinguish incomplete replies and unknown outcomes. Search maintenance releases retired readers, and Working Memory guidance separates confirmed outcomes from future work.
  • When the backend becomes ready after an initial startup error, the App clears the stale error and continues automatically.
  • Full navigation keeps your choice when a slower startup preference read finishes.
  • After a Timeline retry fails, a later retry keeps its busy state until that attempt finishes.
  • With All Spaces selected, new AI Now conversations start in your own default Space. Existing conversations keep their Space.
  • The Skills page background matches the rest of the App in dark mode.
Platform
  • Windows installers update existing installations in place and ask a running App to quit before making changes. Install and uninstall cleanup follows the installed-file manifest, removing obsolete App files while preserving files the App did not install.
  • Linux package updates wait briefly for the backend to finish shutting down and explain how to retry safely when it is still busy.
  • Headless Linux package upgrades stop an active system service before replacing files, then restore it after installation. Services you deliberately stopped stay stopped.
0.10.109Oct 27, 2026

Reliable Recovery, Safer Previews, and Everyday Polish

Keep search rebuilds and AI work recoverable, open previews safely, and retain confirmed settings and edits.

Highlights
  • Search rebuilds on small servers no longer freeze at the first batch. Background index upkeep now shows a waiting state, and a stopped rebuild can be started again after it finishes settling, without restarting the server.
  • AI Now keeps completed tool calls and results after a failed turn. Input refused before sending returns to your draft, while uncertain delivery remains clearly marked.
  • Document previews open in an isolated context. Raw message HTML is displayed as text, and diagram links cannot run scripts in the app.
Improved
  • Access Anywhere presents clearer connection status and actions, with device diagnostics available when you expand them.
  • Profile examples have one consistent preview, support keyboard access, and provide clearer links to Context.
  • Drag the Desktop window by its title strip even with the sidebar drawer open. Window buttons stay clickable, and sidebar labels no longer select text accidentally.
  • Muse and other remote MCP integrations provide one complete, copyable setup prompt using the current workspace address.
  • Menus use consistent text and highlights in light and dark themes. First-run setup retains copy feedback and recovers more reliably from interrupted account changes.
  • Experimental Nowledge Bridge improves recovery of session-receiving receipts and shutdown during preparation, including when storage is full.
Fixed
  • Connection tests correctly route Z.AI and Moonshot Anthropic-compatible endpoints and check JSON mode on Chat Completions routes. Structured background requests through the OpenAI Responses API now send JSON mode correctly.
  • Library Crystal details recover after an interrupted read and refresh when the underlying Memory changes.
  • Cancelling a Timeline review immediately retires its old keyboard actions and returns focus to the original button. Changed targets require a fresh confirmation.
  • Structured background tasks turn reasoning off when supported, or use the lowest listed effort when reasoning cannot be disabled. Your explicit choice takes precedence. For existing providers, open their Settings entry and save once to refresh the effort options.
  • Distillation with remote models reads very long conversations in smaller sections.
  • Background AI tasks use default sampling for models that reject temperature settings. When another model or gateway refuses the setting, later requests adapt for that route and model.
  • Thread fetch, search, and file previews report unavailable or incomplete content instead of showing a stale transcript or partial success. Recovery reasons remain readable in both themes.
  • LAN access and key changes wait for native confirmation. Cancelling leaves settings unchanged, and refreshing Spaces retains confirmed access settings.
  • A failed Memory edit no longer hides accepted label or other Memory updates. Successful detail changes remain visible when returning to the list, and Crystal review stays available in search results.
  • Statistics, activity, Sources, and Memory Health distinguish failed reads from empty data, identify stale snapshots, and offer separate retry actions.
  • Refreshing Memories reruns the submitted search without losing a draft. Saved fields and label changes stay visible while results refresh, including when a refresh fails.
  • Memory and conversation links wait for workspace setup and can be retried after a load failure. Older requests no longer replace the current view after navigation, reconnecting or changing Spaces.
  • AI Now conversations remain usable when tool approvals contain structured query parameters. Approval details and tool inputs can still be inspected as JSON.
  • Codex Desktop tasks marked as agent-created conversations can be saved again. Codex review and helper threads remain excluded, and unknown thread types are identified explicitly.
  • Working Memory refreshes have more room for a complete reply from reasoning models, including Chinese, Japanese and Korean briefings.
  • When final distill review cannot settle one point, only that point is held back while other verified memories are saved. Partial results explain which points remain unsettled.
  • Unrelated websites are blocked from the local Mem backend, Access Anywhere key management and browser control, while Desktop, extension and native tool connections remain supported.
0.10.108Oct 26, 2026

More Reliable Startup, Editing, and AI Settings

Keep drafts and navigation intact, apply model reasoning choices, and recover startup blocked by old installation files.

Highlights
  • On Windows, an unreadable database left by an old installation no longer blocks startup when the current data can be identified safely. The old files are retained, and Mem shows where they were kept.
  • Reasoning effort selected for a Custom OpenAI-compatible model is now saved and sent with requests. OpenCode Go models also show their context window and available reasoning levels in Settings.
  • Skill editors keep your draft through failed saves and detail refreshes. When a save's outcome is uncertain, recovery guidance remains available.
Improved
  • Settings group models and providers under one AI configuration entry, with clearer navigation and consistent neutral styling. Remote API keys stay hidden when copied and are revealed only when you choose.
  • Skill imports, labels, AI Now attachment removal and MCP logs are easier to use from the keyboard. Confirming IME text no longer submits Skill actions or saves unfinished conversation titles.
Fixed
  • Library and Graph keep your latest selection when an older request finishes. Memories refreshes and deletions also respect the current filter and page.
  • Uploads stay with their original connection, failed files remain available for retry, and folder cancellation is respected. An older import preview no longer replaces a newer file selection.
  • Refreshing Library content no longer crashes slide previews, and spreadsheet tabs stay aligned with the sheet being displayed.
  • Access Anywhere clears an old missing-helper error once the helper is available again. If automatic reconnect has paused, Settings explains how to try again.
  • Windows upgrades clean up unused Python-era program files in the installation folder after startup. Memories and settings are preserved.
  • Codex conversations first saved by a CLI older than 0.10.22 can sync new turns again. Existing saved turns keep their original format, and mismatched transcripts are left untouched.
0.10.107Oct 24, 2026

Safer Recovery, Imports, and Everyday Editing

Recover damaged write logs, keep Codex capture moving, and get clearer feedback when imports or saves need attention.

Highlights
  • Mem can recover verified saved changes from damaged write logs left by 0.10.105, including a record that previously caused an allocation crash. Original files are kept, and the recovery notice explains if recent changes may be missing.
  • After updating the CLI, existing Codex sessions can sync again: unchanged sessions are skipped and new turns are appended. If a transcript has diverged, the saved Thread is left intact. Manual saves no longer prevent later automatic capture.
  • Failed backup imports explain whether data may have been partly restored and show available pre-import backup locations. Retrying the same import converges without duplicates. Interrupted processing jobs report possible partial changes after restart instead of remaining busy.
Improved
  • Memories, Threads, conversation imports and Settings have clearer keyboard focus, labels and retry feedback. IME confirmation no longer submits selected actions early, search results stay with the current query, and cancelled plugin settings restore their saved values.
  • Graph controls and filters retain their state as data changes, and topic-loading failures offer recovery actions. Slide editing has clearer keyboard navigation, presentation controls and exports that follow the selected theme.
  • Resume interrupted setup, find favorite conversations in the current Space and browse discovered sessions by project. A searchable AI-tool directory, clearer connection diagnostics and grouped settings make setup and repair actions easier to find.
Fixed
  • Mem can recover verified saved changes from damaged write logs left by 0.10.105, including a record that previously caused an allocation crash. Original files are kept, and the recovery notice explains if recent changes may be missing.
  • Recovery now keeps verified committed changes even when the log contains an unfinished checkpoint. After this recovery, stay on the updated app: older versions cannot open the recovered database. Upgrade again to reopen it.
  • An extra database folder containing only Content no longer traps Windows startup beside the active graph. Mem asks before setting it aside and keeps its files. Cancelling or retrying does not leave startup permanently blocked.
  • After updating the CLI, existing Codex sessions can sync again: unchanged sessions are skipped and new turns are appended. If a transcript has diverged, the saved Thread is left intact. Manual saves no longer prevent later automatic capture.
  • Failed backup imports explain whether data may have been partly restored and show available pre-import backup locations. Retrying the same import converges without duplicates. Interrupted processing jobs report possible partial changes after restart instead of remaining busy.
  • Backup and Skill imports reject conflicting file names before writing and explain what to rename. Skill imports also reject names such as Windows device names. Previously stored Skills remain readable, editable and exportable; devices skip only the affected Skill.
  • Background AI tasks using the built-in Anthropic connection no longer fail because of an unsupported response-format option. Provider connection tests now match the requests those tasks actually send.
  • Thread creation, import, append and detail views keep working with small graph-memory budgets. Saved changes release retained memory before it fills the budget. A backlog from an older version may require a larger budget for one startup.
  • Source saves stay with the document being edited, including long Markdown, even if another document opens before saving finishes. Memory content and labels save together, so a failed save does not leave only part of the edit applied.
  • The launcher copies the latest visible result and runs the current query after edits or a search-mode change. Knowledge extraction reports truncated model responses and retries smaller requests before saving results.
Platform
  • Update dependencies used for PDF parsing, secure connections, Markdown rendering and the browser extension to address published security issues.
0.10.106Oct 23, 2026

Responsive Desktop, Search Repairs, and Clearer Agent Setup

Keep Windows responsive, finish search repairs, and check agent plugins without launching their tools.

Highlights
  • Windows stays responsive while the title bar shows search updates or other busy states, including startup, recovery and shutdown.
  • Settings → Search index → Optimize compacts an index that has slowed after many saves. Search stays available, and the result explains when optimization waits for a rebuild or cannot finish part of the work.
  • The CLI can list the plugin hosts it knows how to inspect and check DeepSeek Harness profile bundles from local files. These checks describe local installation evidence without claiming a plugin is running.
Added
  • AI Now can enable a compiled Skill after a separate confirmation in the conversation, on Desktop and remote self-hosted servers.
  • AI Identity pages show self-reported work separately from confirmed authorship, link writes made with an identity's own key, and offer a way to use that identity in another agent.
  • The CLI can list the plugin hosts it knows how to inspect and check DeepSeek Harness profile bundles from local files. These checks describe local installation evidence without claiming a plugin is running.
  • The CLI can set a model provider request timeout with nmem config provider set --timeout. Editing another provider setting keeps its saved timeout.
Improved
  • Settings → Search index → Optimize compacts an index that has slowed after many saves. Search stays available, and the result explains when optimization waits for a rebuild or cannot finish part of the work.
  • Desktop uses fewer background embedding threads during large syncs. Machines with four or fewer logical CPUs leave one available for other apps, which can make a full search rebuild take longer.
  • A different suggestion about a Memory history decision you already made now waits on Timeline for your review. Keep your decision or use the suggestion; nothing changes until you choose.
  • New conversations captured by the Kimi Code, CodeBuddy, WorkBuddy and Proma plugins can be reviewed for useful memories with your configured AI model. Update the CLI on the capture computer; older saved Threads are not processed again automatically.
Fixed
  • Mem can start and recover verified changes when startup confirms an invalid format in the current write log and the other database locations contain only pre-upgrade data. The recovery notice warns if recent changes may be missing and lists the older databases. Their data is preserved; Mem does not switch to or merge them.
  • Desktop and CLI connection-setting updates are coordinated so overlapping saves preserve unrelated settings. Invalid settings are reported and the original file is retained. When returning to this device, a warning stays visible if CLI settings could not be cleared; Retry preserves the current Desktop connection.
  • Graph tooltips, selection highlights and Details stay in sync after refreshes. Removed items no longer accept actions, while active drags and expanded neighbors are preserved.
  • While a task-model change is pending in the current view, AI Now prevents overlapping changes and keeps progress and result notices in that task and view. A reachable local backend can still receive a change when the browser reports offline, without queuing it for reconnect.
  • Profile loads without waiting for AI identities or rules. On a load failure, Profile and Context offer Retry and enable editing after the profile loads successfully. Name and alias fields also have labels that assistive technology can identify.
  • Large saved changes, such as long Memories, can be recovered after an unexpected shutdown. Damaged write-log records no longer stop recovery with an internal assertion; earlier verified changes are recovered and the original files are retained. Returning to an older version still starts the app, but changes not yet written to the main database may not carry back.
  • Historical database recovery files no longer block startup and remain available for inspection.
  • Memory drafts survive an earlier save finishing after you reopen the form. Edits made while saving and drafts restored after resolving a usage limit stay with the current account.
  • AI Now slide editing keeps text and focus when confirming input-method composition, and editable fields have names that assistive technology can identify.
  • AI Now conversations stay steady during generation when older messages leave the visible list. Hidden file panels stop polling, and earlier copy, save or preview actions cannot affect a newer selection.
  • Timeline answers and Graph Intelligence replies cut off at the model output limit are marked as incomplete, including after reload. The notice suggests a narrower question and explains when no answer was produced.
  • Search repair history keeps its original times across status refreshes. Processing dates follow the app language, and missing repair times are shown as unknown.
  • Switching connections or closing AI Now no longer lets earlier Skill setup results replace the current view's settings. Skills shows discovery progress only after the server accepts a scan.
  • AI Identities stays in its loading state until the current refresh finishes. Results from an earlier refresh cannot overwrite later edits or the list for a newly selected workspace.
  • Background reviews plan within their actual step and reading limits, and Processing history shows their elapsed time.
  • Continue resumes Skill Suggestions from Memories earlier passes already read. A failed continuation keeps the earlier run available to continue.
  • Insight Discovery follows up after reaching its reading limit, and an interrupted follow-up no longer causes the next scheduled review to skip unfinished work.
  • Skills stay in sync when Desktop and the CLI use the same server and key through its base address or /remote-api address. A different workspace still requires an explicit switch.
  • Agent Sessions keeps selections visible across searches, with controls for selected sessions outside the current filter. Available sources stay browsable when another source fails, and large lists load previews on demand.
  • Setup progress animations keep moving smoothly as download status updates arrive.
  • Knowledge graph tasks explain interrupted saves and keep incomplete work from appearing as finished.
  • Windows stays responsive while the title bar shows search updates or other busy states, including startup, recovery and shutdown.
  • On Linux Wayland, a display crash or stalled first window can turn GPU rendering off on the next launch. A notice explains the change, and Settings lets you try it again.
  • After database recovery, confirm that your recent records are complete to finish recovery and remove its temporary files. AI Now can help request cleanup, but still asks for your confirmation.
  • A search repair that is too large to finish automatically now stops with a clear rebuild action instead of keeping large libraries on Updating search indefinitely.
  • Search rebuilds no longer require five percent of an entire large disk to stay free. The safety reserve is capped at 2 GiB, with additional room checked for each write.
  • Merging or removing a Space moves its knowledge-graph entities with its memories. Merge previews show same-name entities that will be combined, and old duplicates left in a removed Space can be merged into the current one.
  • Connected AI tools cannot overturn your reviewed Memory history decisions by marking one Memory as replaced. Changes to an existing decision go through your review.
  • Distill offers Retry only when the run can continue from saved progress. Unusable model answers can be retried up to three times; changed conversations and exhausted retries point to a new Distill request.
  • Ask marks an empty model response as incomplete and keeps the retrieved evidence, instead of presenting an empty or repeated question as the answer.
  • Timeline keeps earlier entries after an interrupted save and reports incomplete saves without pretending they succeeded. Healthy history remains available when another history file is damaged.
  • MCP connection checks and OAuth discovery keep the settings you are currently editing when an earlier check finishes late.
  • Scheduled-task edits stop when their editor closes or changes, including while local authentication is being refreshed.
  • Checking installed plugins no longer launches Codex, Claude Code or Python tools and repeatedly triggers macOS warnings. Uncertain installation or version evidence is shown as unknown.
  • Appending to, renaming, favoriting or updating a Thread respects the personal Space that owns the stored conversation.
  • When a background model response fails after a request may have been sent, Mem keeps unconfirmed usage visible and avoids automatically sending the same work again.
  • Scheduled-task connector actions follow the selected connector and its current settings. Results from an earlier configuration cannot replace the actions you are reviewing.
  • Loading text and AI Now tool, model and Skill icons stay steady during updates. Unfamiliar tool or provider names use a safe fallback icon.
  • Memory, Feed and Thread content stays readable when an imported source has an unfamiliar name.
0.10.105Oct 21, 2026

Steadier Answers, Safer Drafts, and Clearer Recovery

Keep AI answers and draft edits intact, remember Auto and Space choices, and recover from startup trouble with clearer guidance.

Highlights
  • AI Now marks answers cut off at the model's output limit, keeps the notice after a reload, and offers Continue. If the limit is reached before any answer appears, it explains what happened instead of showing a raw provider error.
  • Timeline keeps streaming answers on screen when many response chunks arrive together, without crashing the view.
  • Standalone servers try supported checkpoint recovery automatically. If a graph cannot open, they keep a recovery screen and resume after the file is restored or replaced. The web app warns when recovery may have omitted recent changes.
Added
  • Auto-Sync can save new sessions and the memories learned from them to a Space chosen for their folder or a parent folder. Existing sessions stay where they are. A deleted Space is flagged and new sessions go to Default. Available in personal workspaces.
Fixed
  • AI Now marks answers cut off at the model's output limit, keeps the notice after a reload, and offers Continue. If the limit is reached before any answer appears, it explains what happened instead of showing a raw provider error.
  • Presentation, itinerary, and infographic previews keep the file you selected and newer edits. Reopening refreshes local files, and failed refreshes can be retried.
  • Memory history links keep your choice to retain both memories or keep one as current. Background review cannot rewrite that choice, hide the older Memory, mark it contested again, or copy labels across it.
  • AI Now remembers when you turn Auto off for a task. Closing the app or task, or restarting the server or container, keeps tools asking for approval until you turn Auto on again.
  • AI setup checks saved ChatGPT and GitHub Copilot sign-ins and offers to reconnect when they have expired. A newly configured provider shows as ready only after Mem confirms it was saved and works.
  • Saving, syncing, appending to, editing, or deleting a conversation leaves stored data unchanged when local library identity recovery is required. Deletion also reports the recovery requirement instead of a generic failure.
  • Search starts rebuilding automatically after a recovered local graph becomes active.
  • Background generation and reviews honor the Request timeout saved for your model provider. A request that stalls before answering is reported once instead of being sent again.
  • Distill keeps the messages selected when a run starts, so later additions to a Thread do not invalidate that work. Conversations with messages sharing the same order can also be distilled.
  • When distillation reaches its token budget, including during duplicate review, Processing asks you to raise it. Retry then continues from saved progress. Unknown model-call outcomes can be closed without sending the request again.
  • Desktop keeps specific startup errors and recovery progress visible, and checks that its own backend is ready before opening the workspace.
  • Standalone servers try supported checkpoint recovery automatically. If a graph cannot open, they keep a recovery screen and resume after the file is restored or replaced. The web app warns when recovery may have omitted recent changes.
  • Code blocks without syntax highlighting keep their line breaks and indentation in Threads and memory details. Long lines in Threads scroll within the code block.
  • Connecting to a Mem host where Access Anywhere has never been enabled now explains how to turn it on, instead of suggesting that you wait for startup or restart the host.
  • AI Now offers Retry when Skills, model choices, or schedules cannot load. Skill previews preserve newer draft edits when an earlier save finishes, and an old preview cannot replace current details or feedback.
  • Devin session capture now follows Desktop Auto-Sync, finds Windows history under %APPDATA%, and supports DEVIN_SESSIONS_DIR. Manual imports remain available with nmem t sync --from devin --apply.
  • Search index rebuilding on Windows works when recovered data is stored under a long folder path.
  • Timeline keeps streaming answers on screen when many response chunks arrive together, without crashing the view.
  • On Linux, closing Mem with the window manager, including Super+W or Alt+F4, quits cleanly instead of crashing.
  • AI Now saves or moves memories into the Space you specify. nmem fs write also accepts --space and --space-id; unknown or unwritable destinations are refused instead of silently saving elsewhere.
  • Work with them in AI Now and Open in Graph keep your selection when the destination view has not been opened yet. Switching workspaces or connections cancels an outdated Timeline handoff.
  • Switching accounts keeps sign-in, refresh, and sign-out tied to the current session. Delayed work from a previous account cannot clear a newer session or briefly hide an active Vault license.
  • Context Preview reports clipboard failures, and a missing AI Identity points you to Context to manage identities. AI Identity drafts and character counters also use your selected language.
  • Resizing an AI Now panel stops when you close its preview or cancel the drag, so the next interaction does not keep resizing the panel.
  • Closed or replaced Graph tooltips no longer trigger navigation or close a newer tooltip.
  • AI Now reasoning cards stay readable when a tool returns a thought that is not plain text.
Improved
  • AI Now panels open and close immediately, with keyboard access following what is visible. Progress indicators update without resizing their layout.
  • Automatic Working Memory refresh and Daily Briefing use a character budget that better fits Chinese, Japanese, and Korean, reducing drafts that exceed the size limit.
  • On Windows, the Desktop backend yields CPU to normal-priority applications when the system is busy, while imports and background processing continue.
0.10.104Oct 20, 2026

More Reliable Startup and Model Workflows

Recover supported local startup failures, use Qianwen across model tasks, and keep OCR, long answers, and conversation selections working as expected.

Highlights
  • Desktop can start with a healthy current database even when older graph files remain, and recover supported interrupted checkpoints with a persistent warning when recent changes may be missing.
  • Qianwen AI Platform is available for chat, Vision/OCR, embeddings, and Fast decisions, with chat and decision model choices kept separate.
  • Library image imports recognize the image capabilities of REST-configured OCR models, so supported images produce searchable text without reconfiguring the model in Settings.
Added
  • Qianwen AI Platform is available for chat, Vision/OCR, embeddings, and Fast decisions, with chat and decision model choices kept separate.
  • Library items now offer Copy for AI and Copy ID in a Share menu, with clearer Space details and a confirmation before deletion.
  • Inspect supported local agent plugins with nmem plugins inspect without connecting to a Mem backend. Reports installation and configuration evidence; runtime status is not checked.
Improved
  • Desktop shows the current shutdown phase and avoids walking every search-index file while completing final storage writes.
  • For existing HawDB setups, Processing history distinguishes memory limits, temporary-file limits and the disk-space reserve.
  • Skill suggestions reuse embeddings for unchanged Skills, and duplicate checks keep results from different embedding models or endpoints separate.
  • Background text-search maintenance no longer rewrites the entire index for each batch of changes.
  • Local search-index maintenance releases memory retained by old file versions after they are removed.
Fixed
  • Desktop can start with a healthy current database even when older graph files remain, and recover supported interrupted checkpoints with a persistent warning when recent changes may be missing.
  • Library image imports recognize the image capabilities of REST-configured OCR models, so supported images produce searchable text without reconfiguring the model in Settings.
  • AI Now follows the provider's saved Request timeout while waiting for a slow model to begin answering, instead of stopping after a hidden five-minute limit.
  • Conversation batch selections survive page changes, page actions affect the current page, and mobile lists and pagination stay reachable when selection controls wrap.
  • Profile instructions show a character count and keep over-limit drafts intact instead of silently truncating pasted text.
  • Confirming Chinese or Japanese input no longer submits or closes label entry before composition is complete.
  • ChatGPT subscription setup selects the newest model confirmed by your account while preserving an existing supported selection.
  • Memory index repair verifies date metadata before clearing pending repairs, and progress waits for actual server results before marking items complete.
  • Memory repair status and actions use the selected self-hosted connection. Earlier maintenance requests no longer clear newer selections or refresh an outdated page after you navigate.
  • AI Now’s Auto switch reflects the current task’s confirmed setting. Turning it off stops automatic approval from the next tool onward, including after “Always approve”; tools already approved keep running.
  • Import results, graph-memory errors and Run history explain when the graph has outgrown its startup memory budget, whether a restart is enough, and which setting to change.
  • Local startup no longer skips valid recent graph changes because of a leftover recovery flag.
  • Local Memory Activity now shows the correct latest events and time-window counts, and keeps usage records until their retention cutoff.
  • AI Now approval buttons target the conversation on screen, and copying a Source ID keeps the feedback attached to the copied item.
  • Completed AI Now messages stay visible when scrolling back through conversation history.
  • In shared local workspaces, moves and conversation imports cannot silently move personal items into shared spaces. Refused imports show why, and unknown destinations no longer create empty spaces.
  • Conversation text search stays within the selected Spaces, including when filtering by Source.
  • Local conversation distillation finds the intended conversation and Space, and failure history identifies the cause and affected conversation.
  • Startup recovery can recover verified complete changes saved before a damaged write-log tail, retain the original files, and warn when later changes may be missing.
  • On Windows, recovery after a database write-log error no longer stops on a backup folder left by an earlier import; that folder stays untouched.
  • Knowledge Tree avoids duplicate folder requests when you expand a folder or reopen it while it is still loading.
  • Provider settings preserve unsaved edits through temporary readiness interruptions on the same connection, and refresh the default model after Verify saves configuration.
  • Background model requests that allow retries can fall back to text mode when an OpenAI-compatible provider rejects JSON response mode, without changing saved model settings.
  • Conversation distillation reads complete duplicate-review replies inside code blocks. An unusable received review reply no longer prevents an otherwise valid candidate from being saved as a new Memory.
  • Processing identifies allowance or budget exhaustion during duplicate review and asks for a new Distill when that run cannot continue.
  • Conversation distillation applies memory density when selecting final Memories, so it can consider more useful points while keeping output and context limits.
  • A refused Continue on an outdated distillation run no longer extends the pause on automatic distillation of that conversation.
  • Automatic local graph extraction no longer merges companies solely by an old alias or punctuation-stripped name. Existing graph data still needs separate review.
  • Processing explains when background graph work is paused by a graph-memory or storage problem and shows the next step.
  • Startup recovery can retry a safely interrupted recovery copy while retaining the original files and earlier copies. Repeated interruptions stop further copying with recovery guidance.
  • Graph timeline playback avoids repeated filtering and restores all nodes and edges when its time data or mode resets.
  • AI Now keeps activity expansion and slide edits consistent during rapid updates, including edits made while a slide save is still running.
  • AI Now charts and diagrams keep retired rendering work from changing replacement content, and release render resources when retired.
  • AI Now's writing area adjusts to width changes while keeping long drafts scrollable.

Ready to try Mem?

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

DownloadRead Docs