Faster memory search during embedding provider slowdowns #
Memory search now avoids long retry chains, so queries stay under 600 ms even when an upstream embedding provider slows down.
API changes.
Review suggested memories before saving them #
Review inferred memories from the dashboard and approve, decline, skip, or undo each suggestion before it becomes durable context.
Fixes
- Reconnecting an approved MCP client no longer creates a new API key each time.
- Large Dreaming queues no longer fail during background memory generation.
Forget memories by description with the Memories API #
Use the Memories API to describe what to remove, preview matches with a dry run, then send a second request to forget them.
Fixes
- Stalled or failed Dreaming batches now retry without blocking newer batches in the same container.
Fixes
- Fixed an issue where frequently updated facts were missing from memory profiles.
Search memories by time using natural date phrases #
Search now understands phrases like last week, in March, and between two dates, along with common numeric date formats.
Fixes
- Cloud connections now load and connect correctly through the API.
Install memory in Claude Code, Cursor, OpenCode, and Codex #
The Supermemory CLI can set up memory for Claude Code, Cursor, OpenCode, and Codex from one command.
Improvements
Fixes
- Large PDFs no longer fail to process when text extraction takes longer than usual.
Weekly memory digest emails and history in Nova #
Eligible Nova users receive a Monday recap after weeks with at least three new documents, with digest history and email controls in the web app.
Fixes
- The Memories API now returns custom metadata, and metadata filters work on those values.
- File uploads with invalid metadata now return a validation error instead of a server error.
Fixes
- Documents and memories with numeric metadata now index correctly and remain filterable by number.
- One malformed generated memory no longer fails an entire extraction batch.
Approve MCP connections before they access memory #
Review each MCP client's requested permissions and approve or deny access before it can read or write memory.
Improvements
- Connector sync history now shows which items failed and why.
Run Supermemory locally: self-hosted server for macOS and Linux #
Install the self-hosted server from a published macOS or Linux binary with local authentication and encrypted embedded storage.
Fixes
- The Conversations API now appends turns without replacing, dropping, or duplicating earlier messages.
Fixes
- Rate-limited API requests now return HTTP 429 with Retry-After so clients know when to retry.
Attach files to Nova chats and choose what becomes memory #
Nova keeps attached files with the thread and lets you choose whether each file becomes memory.
Improvements
- Files, links, and notes added from Console now begin memory extraction immediately.
Fixes
- Playground can now save memory guidance for a new container before it has memories.
Export your documents and memories from Settings #
Request a JSON export of up to 25,000 documents and memories. Organization exports require owner or admin access, and download links expire after six hours.
Manage usage, credits, and billing from settings #
Teams can buy more Supermemory usage and manage billing from settings instead of starting a support loop.
Fixes
- Forgotten memories no longer appear in homepage suggestions.
Fixes
- Fixed an issue where MCP connections created a new API key on every request instead of reusing one.
Connect apps, plugins, and knowledge bases from one integrations page #
The integrations catalog brings setup cards, categories, connection state, and deep links into one place.
Fixes
- Fixed an issue where deleting a large organization could fail.
- Fixed an issue where filtering /v3/search by a customId docId returned no results.
Codex plugin routes memories into project Spaces #
Codex can organize captured work into custom memory containers and keep project Spaces named from the context it sees.
Improvements
- Memories now show a temporary title while they are still being processed.
Fixes
- Fixed conversations ingested from plugins being attributed to the API instead of their actual source.
Nova mobile redesign: bottom navigation and app-like layout #
Nova now behaves like a mobile-first app instead of a desktop layout squeezed onto a phone.
Improvements
- Team invites with restricted access now require the Scale plan, with an upgrade option shown on Pro.
Improvements
- Memory generation now uses timestamps and recent history, so new information connects to related memories more accurately.
Fixes
- Document processing now completes even when memory generation fails for part of a batch.
Fixes
- Fixed an issue where Memory of the Day could link to a deleted document and open a blank page.
Fixes
- Large connector syncs now import documents up to your quota instead of failing the whole batch.
- Fixed an issue where email-change confirmation emails sometimes never arrived.
Fixes
- Fixed the memory of the day card opening the list view instead of its source document.
Fixes
- Fixed an issue where Pro plan users were asked to upgrade before inviting their included teammates.
Upload Markdown files in the Console #
Console document uploads now accept Markdown files.
Improvements
Usage-based API billing #
API billing now follows metered usage instead of relying only on fixed quotas.
Improvements
- The file upload endpoint now accepts customId and entityContext, matching the memory add endpoint.
- Usage analytics now report token usage per API key for tracking against plan limits.
Fixes
- Fixed an issue where document metadata updates didn't appear in search results.
Amp plugin: memory that persists across agent runs #
Amp can recall memory at the start of an agent run and save useful outcomes when the run ends.
Fixes
- Fixed /v3/search errors on AND queries with multiple metadata filters.
Fixes
- Deleting an account now removes its related data cleanly.
- MCP clients can now load OAuth discovery metadata reliably.
- Temporary billing-service outages no longer block API requests.
Mount memory as a local folder with SMFS #
SMFS mounts a Supermemory container as a local folder that shells, editors, and agents can use through ordinary file operations.
Fixes
- Fixed PATCH document updates returning a 409 conflict against the document's own filepath.
Improvements
Fixes
- v4 search now reranks results when requested instead of silently returning the original order.
Persistent memory plugin for OpenAI Codex #
Install the Codex plugin to recall project and user memory, capture completed sessions, and explicitly search or save context.
Improvements
- OAuth clients can now discover Supermemory through /.well-known/openid-configuration.
Fixes
- Deleting an organization or account now stops connector syncs and fully disconnects connections.
Choose which Google Drive folders and files sync #
Google Drive sync can be limited to selected folders and files instead of importing everything.
Improvements
Fixes
- MCP connections no longer fail while creating the API key used for a session.
Fixes
- Organizations with overages off now stop at their included allowance instead of recording billable usage.
- Updating a document now counts only newly added content instead of re-counting unchanged text.
Test and tune memory search in the Console #
Run v3 or v4 searches, tune retrieval settings, and inspect scored results, timing, or raw JSON before integrating the API.
Improvements
- Markdown documents now keep headings, tables, and fenced code together when split for indexing.
Fixes
- New API keys now appear immediately in Console, and deleted scoped keys are removed completely.
- Document processing now recovers from blocked generated output instead of failing memory generation.
Polish
- API key previews now preserve the recognizable prefix and use a consistent masked length.
Fixes
- Free organizations no longer keep overage access after a paid plan is canceled.
- The API now rejects connector documents without an access token instead of ingesting their URL as text.
Improvements
- Owners and admins can reset an organization, clearing its data and connections while keeping members, billing, and the default project.
Fixes
- Starting User Insights twice no longer creates competing runs or clears the active report.
Improvements
Fixes
- User Insights now completes for larger organizations and shows its processing state immediately.
- The v4 memory list endpoint now returns results from every requested container tag.
User Insights: reports on user segments and content patterns #
Generate Console reports about user segments, knowledge domains, and content patterns, with an optional prompt to focus the analysis.
Fixes
- You can now forget memories from a Console profile without getting a 404 error.
- Usage emails no longer send the 80% warning after an organization has already reached its limit.
Fixes
- API video ingestion no longer fails while extracting transcripts and descriptions.
- Documents that produce large memory batches now finish processing instead of failing during memory generation.
- Documents no longer fail when a generated memory references another memory that was filtered out.
Fixes
- Image content sent as raw base64 is now recognized as JPEG, PNG, GIF, or WebP instead of failing during summarization.
- Metadata filters containing null characters no longer cause search requests to fail.
Improvements
- Console's API key flow now uses a guided side panel for choosing expiry and scope, then shows the key once with copy and setup links.
- Console's Requests page now summarizes request volume, successful calls, average search latency, and request types, and can filter activity by API key.
Improvements
- During document ingestion, Supermemory can now generate up to three targeted searches to retrieve related existing memories as context.
Fixes
- Documents no longer remain permanently stuck in processing after a workflow stops; Supermemory retries them once and marks a second failure clearly.
- When scheduling a batch document upload fails, only the affected batch is marked failed instead of every queued document.
- API and MCP quota errors now name the limit reached and explain how to continue with overage.
- Truncated model output during ingestion no longer discards every generated memory when only relation data is incomplete.
- Updating a document with the same custom ID no longer creates duplicate memories from facts already extracted.
- Web links that return an authentication, paywall, or rate-limit response during metadata lookup no longer fail the entire import.
- Search requests with an empty query or an out-of-range limit now return a clear client error instead of a server error.
Improvements
Fixes
- Sign-up attempts with an unsupported email domain now return a clear rejection instead of a generic server error.
- Usage limits now account for the full size of each request and allow continued processing when organization overage is enabled.
- Nova no longer intermittently shows an Upgrade to Pro prompt on plugin and connection pages for users with an active paid plan.
- MCP connections authenticated with an API key no longer fail with a missing-organization error.
Fixes
- Account role changes now require administrator permission.
Improvements
- Scoped API keys can now grant access to multiple container tags.
- When removing a connection, you can now keep its imported documents.
Fixes
- Console's API key list no longer crashes when an existing key has no metadata.
- Deleting or merging container tags no longer reports a failure after the operation has already succeeded.
- Google and Microsoft connections now keep syncing when new sign-in credentials replace the old ones.
- Updating only a document's metadata no longer reprocesses unchanged content.
Improvements
Fixes
- Container tags containing colons work again in API requests.
- The Documents API can now retry a failed document with the same custom ID instead of returning the failed record again.
Control API overage at the organization level #
Pro and Scale organizations can choose whether API usage continues into overage from Billing; Enterprise overage remains enabled and cannot be toggled.
Improvements
- Integrators can now disable a scoped API key when a user leaves without deleting that user's memories or container tags.
- Console now lets you choose which GitHub repositories to sync and shows connection details and destinations in sync history.
Fixes
- Container tags with spaces or symbols no longer prevent documents and memories from being indexed.
Improvements
- Plugin sign-in now shows an approval page before returning an API key to the requesting app.
Fixes
- Container tags containing URL-sensitive characters can now be viewed, updated, and deleted correctly in Console.
- Console now shows clear not-found states for missing documents and container tags.
Improvements
Fixes
- Document listing now returns documents matching any selected container tag instead of requiring every tag.
- The API now returns a rate-limit error for throttled keys instead of reporting them as unauthorized.
- v4 profile searches and search requests now record search-query usage correctly.
- Large documents now finish processing instead of failing midway or timing out during summarization.
Search API can return documents, memories, or both #
The v4 Search API can return raw document chunks, extracted memories, or both for each query.
Fixes
- Short notes and questions can now finish processing even when they do not need a generated summary.
Fixes
- Array filters in search, memories, and connections now safely handle user-provided values, and related results stay within the active organization.
Improvements
- Search results now show one copy of identical content, keeping the highest-scoring match.
- Console can create API keys restricted to one container tag, limiting document, search, and profile access to that tag.
- Organization owners and admins can delete a container tag and its associated memories and documents through the API.
Fixes
- Memory extraction now treats dated plans as expiring facts and records why they should expire.
Fixes
- Documents that remain queued now retry automatically up to three times before being marked failed.
Improvements
Fixes
- Appending to an existing document now retains the full combined content and reports its total token count.
Fixes
- Web links that block automated availability checks can now still be imported.
Fixes
- Pages saved by the browser extension now appear as webpages with their original source link instead of plain-text documents.
Fixes
- Reconnecting an approved MCP client now reuses its active API key instead of creating another one.
Fixes
- Fixed an issue that caused some multi-chunk documents to fail during processing.
Fixes
- PDF imports now retry failed pages with a fallback extractor and avoid truncating long documents.
Fixes
- Adding content with an existing custom ID now appends it to that document instead of replacing it.
- Updating a document through the API now replaces and reprocesses the full content instead of appending it.
Improvements
- You can delete the active consumer organization and its stored data from Profile settings.
Fixes
- Updating a document now processes and bills only newly added content instead of unchanged chunks.
Fixes
- MCP and streaming connections now retry brief server overloads before returning a temporary-unavailable error.
Fixes
- Memories added through MCP now count against the memory allowance instead of the API token limit.