Changelog

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

Latest release

0.10.53

Aug 5, 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.

0.10.53

Aug 5, 2026

0.10.53Latest releaseAug 5, 2026

The Shape Of What You Know

Added

The Knowledge Graph has a Types tab. It shows the kinds of things your knowledge is actually made of, and lets you name them, group them, and see how they connect as a diagram. Open it and the canvas becomes a lens on those types; leave and your explore view is exactly as you left it. If you work across languages, one type can carry its names in each of them, so a colleague searching in Chinese and one searching in English land on the same thing.
A conversation can now be handed to another tool by reference. Copy a thread and you get a link the AI can follow back to the full conversation in Mem, instead of a wall of pasted text.

Fixed

Refresh Search no longer spins forever. When a remote provider stopped responding mid-rebuild, the progress indicator had no way to give up and sat there indefinitely. It now reports that the rebuild stalled, and it will not start a second one on top of a worker that has not finished.
Reading text out of images uses a model that can actually see them. Some managed-AI setups from earlier versions were quietly sending pictures to a text-only model, which cannot do the job.
Restoring from a backup brings your AI Now conversations back too. They live outside the knowledge graph, so they used to be left behind when you moved to a new machine. Device paths and credentials stay out of the backup.
Grok models selected from a subscription connect on the first try, and the Library's add menu stays usable while an import is running.
Search is quicker on large libraries, and a run of smaller fixes from recent reports: clearer messages when text extraction has trouble, and steadier behaviour across the app.

0.10.52

Aug 4, 2026

0.10.52Aug 4, 2026

Save The Page, Not Just The Chat

Added

The browser extension now clips ordinary web pages and PDFs, not only AI conversations. Read something worth keeping, save it, and it lands in your Library with the rest of your knowledge. Captures happen on your own machine, and you get a preview of what will be saved before it is.

Fixed

A conversation captured from the browser keeps its middle. Turns could go missing from the middle of a long chat while the beginning and end came across fine, which is the kind of gap you only notice much later, when the reasoning no longer follows.
Answers from reasoning models are no longer cut short. They get the full length you set, and a reply that runs long now has room to finish.
Updating a Memory from an AI tool no longer clears its labels when you were only changing something else.
Restoring settings from a backup works when that backup includes your Spaces.
The Raft connection now lives with the other Connectors instead of off on its own, and the Timeline fills a wide window instead of leaving a column of empty space.

0.10.50

Aug 2, 2026

0.10.50Aug 2, 2026

Straight From Your Reports

Fixed

Connecting Grok through a SuperGrok subscription works again. Mem was sending the model name with a `supergrok/` prefix in front of it, and the service answered that no such model exists, so a perfectly good subscription looked broken. Mem now sends the plain name, and new connections default to Grok 4.5.
Hermes finds Mem again on macOS. A desktop host can start Hermes with a trimmed search path that leaves out the usual places a command line tool installs to, and Hermes would conclude Mem was not there at all. It now looks in those places before giving up.
The Kimi Code plugin no longer ships a hard-coded connection to a local Mem. Run `nmem config mcp show --host kimi-code` and it writes the server block for whichever Mem you actually use, local or remote, so the address and key have one source of truth instead of two that can disagree.
Reading a PDF is faster and cheaper. A page whose only images are small decorative marks no longer triggers a vision pass, so a long text document stops paying for one on every page.
When text extraction on a Library file goes sideways, the notice tells you what actually happened instead of a generic hint. A provider whose only job is embeddings is now labelled that way in Settings, rather than asking you for a default chat model it will never use.
Importing a session from the command line accepts a client name with stray spaces around it, which is how some hosts pass it.

0.10.49

Aug 1, 2026

0.10.49Aug 1, 2026

Proma Conversations Now Come Along

Added

Proma Chat history now has a direct path into Mem. The desktop app can find Proma's local chat conversations, `nmem t sync --from proma --apply` imports the same history from the command line, and automatic session sync can keep those Proma threads updated. Proma Agent sessions still use Proma's lifecycle hooks, so Agent work and regular Chat history each follow the path that matches how Proma records them.
The standalone `nmem` CLI now tells you when a newer version is out, at most once a day, on a single dim line after your command's real output. `nmem status` carries the same information as a field, so an agent can mention it in conversation instead of being nagged on every call. A CLI that came with the desktop app stays quiet, because it updates with the app.

Fixed

Reading text out of an image or a scanned PDF no longer says no vision model is configured when you are signed in to a subscription provider. The credential simply had not been refreshed yet, and Mem now refreshes it before it looks.
If you point Mem at your own data directory, it stays there. An empty directory in a non-standard location used to look like a misplaced install, so Mem would quietly repoint itself at the graph in the standard location instead.
Saving one section of Settings no longer writes back a stale copy of another. Your profile, your agent profiles, and your rules are only written by the screen that owns them.
A backup taken while Mem is running now skips the database's temporary lock and in-flight checkpoint files instead of copying them into the archive.
Terminal tools in AI Now now run with your login shell environment, so anything you installed through your shell profile is on PATH where the agent expects to find it. Sending feedback from the app also attaches a bounded slice of recent server logs, so a report arrives with the evidence already in it.
In Files, a memory path under `by-id` is an exact lookup again rather than a partial match on the id, so a shortened id cannot land you on a different memory. An Agent identity with an unusable id is now refused up front instead of creating a profile that nothing can reach.
Proma setup instructions now include the `enabled: true` field required by recent Proma builds before the MCP server and related hooks are treated as active.

0.10.48

Jul 30, 2026

0.10.48Jul 30, 2026

What A Real Corpus Reveals

Fixed

Code blocks are highlighted again on the web, and the Feedback button no longer swallows your click. Both were found by walking every surface with a real 78 MB corpus behind it, and both were invisible against an empty workspace. The Feedback button was waiting on a script that never loads, so the click queued against an event that would never fire; it now has a deadline and falls back to the same place the widget would have filed to.
Task warnings in the Timeline clear themselves once the same kind of task succeeds again, so a one-off hiccup stops sitting there looking like an ongoing problem. The history stays in the journal.
Hosted workspaces now get skill rot detection. Cloud never ran the curator, so a hosted user's skills could quietly stop working with nothing to say so.
Connecting a Cloud account from the desktop app is wired end to end, including getting your first access key.
AI Now presentations use a refreshed bento slide format. Remote embedding retries when a provider returns an error inside an otherwise successful response, `nmem models status` reports the live search index state, and `nmem` writes a working MCP host by default.

0.10.47

Jul 30, 2026

0.10.47Jul 30, 2026

Filters That Tell The Truth

Fixed

Filtering Memories by label works again, and an unknown label now says so instead of quietly returning everything. This applies everywhere the filter is offered: search, list, the API, and `nmem memories search -l` / `nmem memories list -l`. A filter that silently ignores you is worse than one that fails, because the full list looks like a real answer.
The CLI now tells you when the embedding runtime is degraded. `nmem status` and `nmem models status` read the live runtime state and warn, and the JSON output carries it too, so a script can notice as well. Previously meaning-based search could be quietly running on a fallback with nothing on screen to explain it.
The local embedding model now loads when it is first needed rather than at startup, so launching is quicker, and remote embedding calls time out instead of hanging.
Sharpening a skill stays within bounds, and you can replace a skill's body directly instead of only nudging it with instructions.

0.10.46

Jul 29, 2026

0.10.46Jul 29, 2026

Steady Background Work

Managed AI keeps working through a token rotation, search repair finishes the exact work it started, the Timeline shows a dependable history, and Stop actually stops a running review.

Fixed

Managed AI now keeps working when your access token rotates. Chat and embeddings previously kept using the token they started with, so roughly once an hour a request could fail even though your account was healthy. A rejected request is now retried once with the fresh token, and long-running consumers like embeddings and the browser extension pick up the new credential without a restart.
Search Index repair now finishes exactly the work it started. Targeted Memory repair is back, repair scope survives a restart, and a rebuild no longer restarts itself in a loop. Non-semantic edits keep their existing Cloud embeddings instead of paying for them again.
Stop now stops a review that is already running, instead of only removing it from the queue.
Timeline history is dependable again: the calendar shows the right day, long histories page smoothly, and re-uploading the same file keeps it as context instead of dropping it.
Sending a bundle to Cloud no longer fails as a whole because of a few things Cloud cannot hold yet. Community groupings and the links from memories back to their source passages stay on this device and are named for you, while the rest of the bundle lands.
In Team, the Raft entry opens on the first click, and handing off a hosted Raft connection or rotating its access key works again.
PDFs added from the Timeline are processed reliably, stopped tasks read as a neutral state rather than a failure, and the ChatGPT connector link points at the current setup page.

0.10.45

Jul 28, 2026

0.10.45Jul 28, 2026

Search and Review Reliability

Search repair now works with more embedding providers, conversation processing respects your choices, updated Apps reliably use the matching Mem server, and exported knowledge now records how far each piece has been reviewed.

Added

The Library export now follows Open Knowledge Format v0.2, the version Google published in July, so bundles record how far each piece of knowledge has been checked and not just what you know. Crystals you confirmed are marked as reviewed by a person and appear at the highest trust tier in any OKF-aware reader, while anything the agent proposed without your sign-off stays explicitly unverified. Dismissed, archived, and replaced entries are marked as no longer current instead of sitting alongside live knowledge, and sources are attributed at the exact sentence they support rather than listed at the bottom. The reviewed-by-a-person mark is deliberately anonymous, so sharing a bundle never exports your identity.

Fixed

Command-line tools installed with Home Manager, Nix, Homebrew, or another package manager now remain under that manager's control. If you turn off the App-managed CLI, updates no longer bring back the Terminal setup reminder.
The desktop App now verifies the bundled Mem server before starting it. An older server left by a previous installation can no longer hide new Settings or Memory Health controls after the App has been updated.
Opening Mem from another device on your LAN now shows the self-hosted API-key connection instead of asking you to sign in to Nowledge Cloud. Cloud sign-in remains isolated from local access keys.
Browser login links created by `nmem key --show-login` are now single-use and expire after one minute. Rotating your access key also cancels any unused links.
Search Index rebuilds now complete with Gemini and more OpenAI-compatible embedding services. If a provider returns an incomplete or conflicting result, Mem still keeps your previous searchable index intact.
Memory saves and edits remain responsive while a long index write finishes safely in the background. Moves, lifecycle changes, and deletions also stay aligned with Search without reporting completed work as failed or starting duplicate repairs.
Search refresh now settles cleanly when recent changes finish indexing in the background. Settings shows that work as an update in progress instead of asking you to rebuild the same index again.
Restoring a backup now schedules both Memory and conversation search to refresh from the restored data, so imported messages return to search during recovery.
Turning off automatic conversation processing now remains authoritative even for work queued by an older build or while another Settings page is being saved. Completed processing also returns to Timeline as a visible audit entry.
Memory relationship review now keeps every evidence-backed decision from its bounded candidate set instead of dropping later valid matches. Deleted or moved Memories no longer leave an impossible review card, and older cards open in the correct Space when that can be determined safely.
When a Space merge cannot safely determine where a conversation belongs, Settings now takes you directly to the conversations that need review instead of leaving the merge blocked without a next step.
Background relationship maintenance now finishes each review batch before making room for other Processing work. If a model or search service is unavailable, Mem pauses that batch and retries later without holding up the queue.
Scanned PDFs now process one safely sized page at a time for remote vision while keeping the higher-quality local-model path. Large documents no longer build every page image at once, and unfinished OCR is shown as partial or ready to retry instead of complete.
Menus and pickers remain clearly readable in dark mode on Linux, including Library import and export actions.

0.10.44

Jul 27, 2026

0.10.44Jul 27, 2026

Clearer Connections and Agent Identity

Agent identities are easier to understand, first-time connection choices are clearer, and the App stays composed across loading and smaller windows.

Improved

Team Settings now gives Agent identities their own focused section, so you can see and manage who contributed knowledge without navigating a large integration card.
First launch now explains the available connection paths more clearly and keeps imported knowledge visible as you finish setup.
Sidebar actions stay within the window at smaller sizes, and lazily loaded views keep a visible transition instead of briefly appearing blank.

0.10.43

Jul 27, 2026

0.10.43Jul 27, 2026

Remote Search Repair

Search Index rebuilds now work with more OpenAI-compatible embedding services, and Thread search now recovers more smoothly from temporary degradation.

Fixed

Search Index rebuilds now work with remote embedding gateways that label every result as index 0 while returning the vectors in request order. Mem still rejects genuinely incomplete or ambiguous responses, and a failed rebuild continues to preserve the previous searchable index.
Thread search now preserves full message results more reliably and recovers more smoothly from temporary degradation.

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

0.10.53

Aug 5, 2026

0.10 is a ground-up engine rewrite

0.10.53

Aug 5, 2026

0.10.53Latest releaseAug 5, 2026

The Shape Of What You Know

Added

The Knowledge Graph has a Types tab. It shows the kinds of things your knowledge is actually made of, and lets you name them, group them, and see how they connect as a diagram. Open it and the canvas becomes a lens on those types; leave and your explore view is exactly as you left it. If you work across languages, one type can carry its names in each of them, so a colleague searching in Chinese and one searching in English land on the same thing.
A conversation can now be handed to another tool by reference. Copy a thread and you get a link the AI can follow back to the full conversation in Mem, instead of a wall of pasted text.

Fixed

Refresh Search no longer spins forever. When a remote provider stopped responding mid-rebuild, the progress indicator had no way to give up and sat there indefinitely. It now reports that the rebuild stalled, and it will not start a second one on top of a worker that has not finished.
Reading text out of images uses a model that can actually see them. Some managed-AI setups from earlier versions were quietly sending pictures to a text-only model, which cannot do the job.
Restoring from a backup brings your AI Now conversations back too. They live outside the knowledge graph, so they used to be left behind when you moved to a new machine. Device paths and credentials stay out of the backup.
Grok models selected from a subscription connect on the first try, and the Library's add menu stays usable while an import is running.
Search is quicker on large libraries, and a run of smaller fixes from recent reports: clearer messages when text extraction has trouble, and steadier behaviour across the app.

0.10.52

Aug 4, 2026

0.10.52Aug 4, 2026

Save The Page, Not Just The Chat

Added

The browser extension now clips ordinary web pages and PDFs, not only AI conversations. Read something worth keeping, save it, and it lands in your Library with the rest of your knowledge. Captures happen on your own machine, and you get a preview of what will be saved before it is.

Fixed

A conversation captured from the browser keeps its middle. Turns could go missing from the middle of a long chat while the beginning and end came across fine, which is the kind of gap you only notice much later, when the reasoning no longer follows.
Answers from reasoning models are no longer cut short. They get the full length you set, and a reply that runs long now has room to finish.
Updating a Memory from an AI tool no longer clears its labels when you were only changing something else.
Restoring settings from a backup works when that backup includes your Spaces.
The Raft connection now lives with the other Connectors instead of off on its own, and the Timeline fills a wide window instead of leaving a column of empty space.

0.10.50

Aug 2, 2026

0.10.50Aug 2, 2026

Straight From Your Reports

Fixed

Connecting Grok through a SuperGrok subscription works again. Mem was sending the model name with a `supergrok/` prefix in front of it, and the service answered that no such model exists, so a perfectly good subscription looked broken. Mem now sends the plain name, and new connections default to Grok 4.5.
Hermes finds Mem again on macOS. A desktop host can start Hermes with a trimmed search path that leaves out the usual places a command line tool installs to, and Hermes would conclude Mem was not there at all. It now looks in those places before giving up.
The Kimi Code plugin no longer ships a hard-coded connection to a local Mem. Run `nmem config mcp show --host kimi-code` and it writes the server block for whichever Mem you actually use, local or remote, so the address and key have one source of truth instead of two that can disagree.
Reading a PDF is faster and cheaper. A page whose only images are small decorative marks no longer triggers a vision pass, so a long text document stops paying for one on every page.
When text extraction on a Library file goes sideways, the notice tells you what actually happened instead of a generic hint. A provider whose only job is embeddings is now labelled that way in Settings, rather than asking you for a default chat model it will never use.
Importing a session from the command line accepts a client name with stray spaces around it, which is how some hosts pass it.

0.10.49

Aug 1, 2026

0.10.49Aug 1, 2026

Proma Conversations Now Come Along

Added

Proma Chat history now has a direct path into Mem. The desktop app can find Proma's local chat conversations, `nmem t sync --from proma --apply` imports the same history from the command line, and automatic session sync can keep those Proma threads updated. Proma Agent sessions still use Proma's lifecycle hooks, so Agent work and regular Chat history each follow the path that matches how Proma records them.
The standalone `nmem` CLI now tells you when a newer version is out, at most once a day, on a single dim line after your command's real output. `nmem status` carries the same information as a field, so an agent can mention it in conversation instead of being nagged on every call. A CLI that came with the desktop app stays quiet, because it updates with the app.

Fixed

Reading text out of an image or a scanned PDF no longer says no vision model is configured when you are signed in to a subscription provider. The credential simply had not been refreshed yet, and Mem now refreshes it before it looks.
If you point Mem at your own data directory, it stays there. An empty directory in a non-standard location used to look like a misplaced install, so Mem would quietly repoint itself at the graph in the standard location instead.
Saving one section of Settings no longer writes back a stale copy of another. Your profile, your agent profiles, and your rules are only written by the screen that owns them.
A backup taken while Mem is running now skips the database's temporary lock and in-flight checkpoint files instead of copying them into the archive.
Terminal tools in AI Now now run with your login shell environment, so anything you installed through your shell profile is on PATH where the agent expects to find it. Sending feedback from the app also attaches a bounded slice of recent server logs, so a report arrives with the evidence already in it.
In Files, a memory path under `by-id` is an exact lookup again rather than a partial match on the id, so a shortened id cannot land you on a different memory. An Agent identity with an unusable id is now refused up front instead of creating a profile that nothing can reach.
Proma setup instructions now include the `enabled: true` field required by recent Proma builds before the MCP server and related hooks are treated as active.

0.10.48

Jul 30, 2026

0.10.48Jul 30, 2026

What A Real Corpus Reveals

Fixed

Code blocks are highlighted again on the web, and the Feedback button no longer swallows your click. Both were found by walking every surface with a real 78 MB corpus behind it, and both were invisible against an empty workspace. The Feedback button was waiting on a script that never loads, so the click queued against an event that would never fire; it now has a deadline and falls back to the same place the widget would have filed to.
Task warnings in the Timeline clear themselves once the same kind of task succeeds again, so a one-off hiccup stops sitting there looking like an ongoing problem. The history stays in the journal.
Hosted workspaces now get skill rot detection. Cloud never ran the curator, so a hosted user's skills could quietly stop working with nothing to say so.
Connecting a Cloud account from the desktop app is wired end to end, including getting your first access key.
AI Now presentations use a refreshed bento slide format. Remote embedding retries when a provider returns an error inside an otherwise successful response, `nmem models status` reports the live search index state, and `nmem` writes a working MCP host by default.

0.10.47

Jul 30, 2026

0.10.47Jul 30, 2026

Filters That Tell The Truth

Fixed

Filtering Memories by label works again, and an unknown label now says so instead of quietly returning everything. This applies everywhere the filter is offered: search, list, the API, and `nmem memories search -l` / `nmem memories list -l`. A filter that silently ignores you is worse than one that fails, because the full list looks like a real answer.
The CLI now tells you when the embedding runtime is degraded. `nmem status` and `nmem models status` read the live runtime state and warn, and the JSON output carries it too, so a script can notice as well. Previously meaning-based search could be quietly running on a fallback with nothing on screen to explain it.
The local embedding model now loads when it is first needed rather than at startup, so launching is quicker, and remote embedding calls time out instead of hanging.
Sharpening a skill stays within bounds, and you can replace a skill's body directly instead of only nudging it with instructions.

0.10.46

Jul 29, 2026

0.10.46Jul 29, 2026

Steady Background Work

Managed AI keeps working through a token rotation, search repair finishes the exact work it started, the Timeline shows a dependable history, and Stop actually stops a running review.

Fixed

Managed AI now keeps working when your access token rotates. Chat and embeddings previously kept using the token they started with, so roughly once an hour a request could fail even though your account was healthy. A rejected request is now retried once with the fresh token, and long-running consumers like embeddings and the browser extension pick up the new credential without a restart.
Search Index repair now finishes exactly the work it started. Targeted Memory repair is back, repair scope survives a restart, and a rebuild no longer restarts itself in a loop. Non-semantic edits keep their existing Cloud embeddings instead of paying for them again.
Stop now stops a review that is already running, instead of only removing it from the queue.
Timeline history is dependable again: the calendar shows the right day, long histories page smoothly, and re-uploading the same file keeps it as context instead of dropping it.
Sending a bundle to Cloud no longer fails as a whole because of a few things Cloud cannot hold yet. Community groupings and the links from memories back to their source passages stay on this device and are named for you, while the rest of the bundle lands.
In Team, the Raft entry opens on the first click, and handing off a hosted Raft connection or rotating its access key works again.
PDFs added from the Timeline are processed reliably, stopped tasks read as a neutral state rather than a failure, and the ChatGPT connector link points at the current setup page.

0.10.45

Jul 28, 2026

0.10.45Jul 28, 2026

Search and Review Reliability

Added

The Library export now follows Open Knowledge Format v0.2, the version Google published in July, so bundles record how far each piece of knowledge has been checked and not just what you know. Crystals you confirmed are marked as reviewed by a person and appear at the highest trust tier in any OKF-aware reader, while anything the agent proposed without your sign-off stays explicitly unverified. Dismissed, archived, and replaced entries are marked as no longer current instead of sitting alongside live knowledge, and sources are attributed at the exact sentence they support rather than listed at the bottom. The reviewed-by-a-person mark is deliberately anonymous, so sharing a bundle never exports your identity.

Fixed

Command-line tools installed with Home Manager, Nix, Homebrew, or another package manager now remain under that manager's control. If you turn off the App-managed CLI, updates no longer bring back the Terminal setup reminder.
The desktop App now verifies the bundled Mem server before starting it. An older server left by a previous installation can no longer hide new Settings or Memory Health controls after the App has been updated.
Opening Mem from another device on your LAN now shows the self-hosted API-key connection instead of asking you to sign in to Nowledge Cloud. Cloud sign-in remains isolated from local access keys.
Browser login links created by `nmem key --show-login` are now single-use and expire after one minute. Rotating your access key also cancels any unused links.
Search Index rebuilds now complete with Gemini and more OpenAI-compatible embedding services. If a provider returns an incomplete or conflicting result, Mem still keeps your previous searchable index intact.
Memory saves and edits remain responsive while a long index write finishes safely in the background. Moves, lifecycle changes, and deletions also stay aligned with Search without reporting completed work as failed or starting duplicate repairs.
Search refresh now settles cleanly when recent changes finish indexing in the background. Settings shows that work as an update in progress instead of asking you to rebuild the same index again.
Restoring a backup now schedules both Memory and conversation search to refresh from the restored data, so imported messages return to search during recovery.
Turning off automatic conversation processing now remains authoritative even for work queued by an older build or while another Settings page is being saved. Completed processing also returns to Timeline as a visible audit entry.
Memory relationship review now keeps every evidence-backed decision from its bounded candidate set instead of dropping later valid matches. Deleted or moved Memories no longer leave an impossible review card, and older cards open in the correct Space when that can be determined safely.
When a Space merge cannot safely determine where a conversation belongs, Settings now takes you directly to the conversations that need review instead of leaving the merge blocked without a next step.
Background relationship maintenance now finishes each review batch before making room for other Processing work. If a model or search service is unavailable, Mem pauses that batch and retries later without holding up the queue.
Scanned PDFs now process one safely sized page at a time for remote vision while keeping the higher-quality local-model path. Large documents no longer build every page image at once, and unfinished OCR is shown as partial or ready to retry instead of complete.
Menus and pickers remain clearly readable in dark mode on Linux, including Library import and export actions.

0.10.44

Jul 27, 2026

0.10.44Jul 27, 2026

Clearer Connections and Agent Identity

Agent identities are easier to understand, first-time connection choices are clearer, and the App stays composed across loading and smaller windows.

Improved

Team Settings now gives Agent identities their own focused section, so you can see and manage who contributed knowledge without navigating a large integration card.
First launch now explains the available connection paths more clearly and keeps imported knowledge visible as you finish setup.
Sidebar actions stay within the window at smaller sizes, and lazily loaded views keep a visible transition instead of briefly appearing blank.

0.10.43

Jul 27, 2026

0.10.43Jul 27, 2026

Remote Search Repair

Search Index rebuilds now work with more OpenAI-compatible embedding services, and Thread search now recovers more smoothly from temporary degradation.

Fixed

Search Index rebuilds now work with remote embedding gateways that label every result as index 0 while returning the vectors in request order. Mem still rejects genuinely incomplete or ambiguous responses, and a failed rebuild continues to preserve the previous searchable index.
Thread search now preserves full message results more reliably and recovers more smoothly from temporary degradation.

Ready to try Mem?

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

Download Read Docs