Changelog

Forget

• New scopes: actions, kv, and events. The default is still ["memories"], so to forget something everywhere, list all six.
• The result now has not_searched, the scopes it didn't look in, and truncated, which is true when there's more to do. Check both before saying something is gone.
• Actions are redacted, not deleted, so a denied action can't be proposed and run again. Any that could still run become superseded. executing_actions lists those already running.
• Phrase matching now ignores case and accents in any script, and it covers journal source_refs and fact tags and source_ref.

New

• POST /v1/space/key replaces a space's key. The old key stops working at once, and everything in the space stays, including a stored OpenAI key. REST only.
• PATCH /v1/actions/{id} merges into an action's meta, in any status. Also the action_update MCP tool and the action.update batch op. The MCP server now has 34 tools.
• search.indexing on recall counts the matching facts that may not be findable by meaning yet.
• open_records_total and open_limit on GET /v1/context.
• error.details.op_index and op on a failed batch, including failed preconditions.

Fixed

• A q with no letters or digits, such as ?!, now finds nothing. It used to list every fact.
• OpenAPI: created is a plain boolean on both responses, weekly is nullable in a form client generators keep, the name pattern no longer has a stray /i, and the batch limit says 50, the real limit. If you generate a client from the document, regenerate it.

Unchanged on purpose

At launch

• Spaces with anonymous keys: POST /v1/spaces returns a key with no signup.
• Facts with full-text recall.
• A journal, records with compare-and-set, an approval ledger with idempotent proposals, KV, and capped events.
• GET /v1/context loads what matters at the start of a task, and POST /v1/batch applies writes atomically.
• forget with content-free tombstones and journal redaction.
• Dreaming: nightly housekeeping, a model stage on your agent's own model or on Workers AI, safety rules for what you stated, and a 7-day undo.
• A REST API with an OpenAPI 3.0 document, and a stateless MCP server.
• Docs and agent guides, with every page also served as Markdown, plus llms.txt, llms-full.txt, and an agent skill.

Added before 0.2.0

• Meaning search: recall matches by meaning as well as by words, within a daily budget.
• Your own OpenAI key for embeddings, with PUT /v1/space/embeddings. It's stored encrypted and is write-only and REST-only.
• Recall lifts facts about the person a question names, and matches other word forms.
• extract_job and extract_submit let an agent go through a conversation with its own model. The server runs no model and keeps no copy of the conversation.
• A retrieval benchmark on LongMemEval.
• MCP lists four tools by default: get_context, remember, recall, and forget. Add ?tools=all to list the rest. Every tool can still be called by name.
• recall with no question, or GET /v1/memories, reaches every fact. It returns total and next_cursor, and doesn't count as a recall.
• MCP gained 16 tools, for 33 in all. Creating or deleting a space and managing an OpenAI key stay REST-only.
• The API reference lets you try calls from the browser. CORS allows only openmemory.sh and localhost, and your key goes only to the API. api.openmemory.sh/docs now redirects to the reference.