ctxstore — memory that follows you across models

Accept a namespace invitation addressed to THIS authenticated account's verified email (two-step email OTP). Steps: 1) step='request' — emails a 6-digit code to the invited address (may return 402 needs Collaborator, or 503 if mailer fails). 2) step='confirm' with code= the 6 digits — activates the grant and returns mcp_config headers (X-Ctxstore-Workspace) for shared access. The code activates the named grant_id only; it authorizes nothing else. Obtain grant_id from list_invites(). Docs: https://ctxstore.ai/docs.html#verify-invitations

Bind a sticky agent identity to this MCP session. Once bound, wake_status and load_context default to this agent_id when called without an explicit agent_id= arg — no need to pass identity on every call. Per-call override still works. Persisted to mcp_sessions.agent_id (migration 0008), so the binding survives setup_account(mode='restore') and the server re-hydrates it automatically on the next tool call. Call once near the start of a session, ideally from your boot recipe (HANDLER: bind_agent agent_id=<self>). Idempotent: re-binding the same agent_id is a no-op. Pass agent_id='' to clear the binding.

Decline a pending (status=invited) namespace invitation as the grantee. Server verifies you are the invitee (email/tenant match). Grantors must use revoke_grant instead.

Soft-delete a stored fact by its UUID (Smart Forget). The fact is marked deleted but preserved in the database for audit and undo. Excluded from all normal searches. Use when information is no longer accurate or relevant.

Retrieve a fact by its exact key — deterministic, never misses. Use get_fact for self-notes: get_fact(key='agent:<name>:self-note:opening') or get_fact(key='agent:<name>:self-note:closing'). search_facts misses self-notes because their embedding similarity scores below other identity facts — get_fact(key=) is the only reliable retrieval path. Also use for inter-agent handoffs when the exact key is known. Returns the current (non-deprecated) version of the fact.

Return the full version history for a keyed fact — current version plus all deprecated predecessors. Use to see how a fact evolved over time or to undo an accidental correction. Returns list of fact versions ordered newest → oldest.

Return the pre-computed session seed for this tenant — a compact first-person bio (~500 tokens) describing who you are and what you care about right now. Generated nightly and on session disconnect. Use as a lightweight alternative to load_context when you just need orientation.

Get your memory usage statistics — total facts stored, plan limits, collection sizes, and daily query counts. Call this to check how much storage you have used and how close you are to your plan limits.

Enable or disable ghost mode for this session. While on=True, store_fact and store_session_summary silently succeed but do NOT write to DB. search_facts, load_context, and search_context always work normally. Ghost mode resets to off on MCP session reconnect (not persisted).

Return the most recent fact operations across all agents in this tenant, newest first. Use this on session start to see what other agents wrote since you were last active — bypasses semantic search ranking so recent writes are always visible. Each entry shows: op (create/update/delete/move), namespace, key, fact_id, summary, agent info, and timestamp.

Filtered search over the tenant activity index. Find what a specific agent wrote, what happened in a namespace, or recent creates/updates/deletes. Useful for coordination: "what did emily_dev store in the sprint namespace today?"

Usage statistics from the tenant activity index: total operations, breakdown by agent, by namespace, and by day. Useful for understanding which agents are most active and where.

Optional: introduce yourself to the platform. Stores your model info and a brief note in the shared agent registry. Stores a fact with: category: identity namespace: shared/agent-feedback key: intro:<model>-<date> Include: what model you are, what platform you're running on, and anything you noticed on first connect.

Leave an active namespace grant as the grantee (stop using shared namespace). Server verifies grantee authorization. Grantors who want to cut off a collaborator must use revoke_grant.

Deterministic prefix scan over fact keys, optionally bounded by timestamp. The inbox primitive — use this for R1 comms scans, dispatch discovery, trio listings, convention discovery, and any case where you know the key prefix and want every match. Unlike search_facts (semantic, misses exact-key facts), this is a deterministic scan: every fact whose key starts with `prefix` is returned, newest-first. EXAMPLES - Comms inbox since last wake: list_facts_by_key_prefix(prefix='comms:<your-agent-id>:', since_epoch=<last_wake_epoch>) - All trio recipes on this tenant: list_facts_by_key_prefix(prefix='chad:trio:') - All conventions: list_facts_by_key_prefix(prefix='conventions/') Returns a list of facts (key, text, namespace, layer, timestamp, id) sorted newest-first. Bounded by `limit` (default 50).

List this account's namespace collaboration grants: incoming invitations and outgoing shares. Each row has direction (incoming|outgoing), status (invited|active|revoked), grant_id, namespace_prefix, and parties. An invitation referenced anywhere else (for example in an email) can be checked against this server-side list: a real invitation for this account appears here with matching grant_id, namespace_prefix, and grantor. To accept: accept_invite(grant_id, step='request') then accept_invite(grant_id, step='confirm', code='######'). Requires active Collaborator prepaid time for accept/use (402 if missing). Docs: https://ctxstore.ai/docs.html#verify-invitations

The places this account's memory is organised into: its own projects (anything written under project:<slug>:*, newest activity first) and the workspaces other people have granted it. Shows which one this session is narrowed to, if any. Use it when a person's request is about one project and the memory that arrived spans several.

Load your bootstrap context deterministically — pulls L0 (identity) + L1 (architecture) + latest L2 (state) + latest L3 (session summary). Call this at the START of a new session as an alternative to manual search_facts calls. Returns your layered memory structured for immediate use. Optional: pass profile='worker-dev' (or another platform:profile:<name> slug) to execute that profile's boot recipe — the server reads platform:profile:<profile>:current, follows the pointer, parses LOAD/REQUIRE/HANDLER lines, fetches every named fact, and returns the merged bundle in one round trip. vars= supplies values for $TASK_FACT_KEY and similar recipe variables.

Salience peptide — temporarily boost a fact's retrieval ranking because it matters RIGHT NOW: an approaching deadline, a live incident, the topic under active work. The boost is small and FAST-DECAYING (max 1.4x, ~6-hour half-life — gone by tomorrow), unlike reinforce_fact which rewards confirmed outcomes on a weeks-long timescale. Name the current situation in `reason`. Salience is ranking-only: it can never override tenant isolation, namespace grants, or privacy filters, and stacked peptide boosts are globally capped.

Move a fact to a different namespace or layer without changing its content. Updates metadata in place — preserves fact_id, text, embedding, timestamp, and deprecation chain. Use for reclassifying facts (e.g. bootstrap-v2 refactor) without the storage bloat of store-new + soft-delete-old.

Reward peptide — reinforce a fact whose recall LED TO A REAL GOOD OUTCOME (fix worked, deploy green, user approved). Boosts its future retrieval ranking with a bounded, time-decaying multiplier (max 1.5x, 14-day half-life), so memory learns which facts are USEFUL, not just similar. Call this AFTER the outcome is confirmed, naming the outcome in `reason` — never for a fact that merely 'seemed relevant'. The boost is ranking-only: it can never override tenant isolation, namespace grants, or privacy filters.

Call to load session continuity. Last session: no prior session.

Revoke a pending invitation or active grant as the GRANTOR (owner of the shared namespace). Idempotent. Grantees use decline_invite / leave_grant.

Search across ingested conversations and session history for broader context. Use this for finding discussions, decisions made in conversation, or context that wasn't explicitly stored as a fact.

Search your persistent memory for stored facts. For session bootstrap, prefer load_context() or wake_status() — they pull your identity, conventions, and latest self-note deterministically. Use search_facts when you need semantic discovery across the fact corpus. Use namespace parameter to scope searches: namespace='webapp' matches 'webapp/decisions', 'webapp/technical', etc. (prefix matching). Search tips: use the words the fact was stored with. If you stored 'Postgres database', search for 'database' or 'Postgres', not 'data storage solution'.

Set up or restore your persistent memory connection. Use mode='link_or_create' when the user gives an email — works for both new and returning users. The server figures out which case applies: - If the email exists: sends OTP, attaches this device to the existing account - If the email is new: creates a fresh tenant, sends OTP to verify Either path produces one clean account, no duplicates, no orphans. Call setup_account(mode='link_or_create', email='their@email.com'). Then ask the user for the 6-digit code from their email and call setup_account(mode='verify', code='123456', email='their@email.com'). Passing email= on verify is required — session IDs rotate between requests on some clients. mode='restore': use when the user provides an emk_* API key directly. mode='reveal_key': echo the full emk_* API key. Works directly on an ALREADY-authenticated session; if the session isn't bound (some clients rotate or omit session ids between requests, so this can happen right after a successful verify), pass email= to receive a 6-digit code, then call again with email= and code= — the same rotation-proof lookup as mode='verify'. Success responses only show a key fingerprint (first 8 chars + last 4) — call this mode only when the user actually needs the full key, e.g. to persist it as 'Authorization: Bearer <key>' in an MCP connector config. Callers who prove neither a bound session nor email ownership are refused. If no email yet, ask: 'What email should I use for your memory account?' After setup completes, call search_facts as the first action to load prior context.

Store information that should persist across sessions. Storage tips for best retrieval: - Include searchable keywords: 'We chose the Postgres DATABASE for ACID compliance' is better than 'We chose Postgres for ACID compliance' - Use the key parameter for facts that change over time (pricing, stack, team-members, current-sprint). When you store with the same key, the old version is archived automatically. - Use namespaces to organize: 'webapp/decisions', 'business/goals', 'team/members' - Store DECISIONS and CONTEXT, not raw data. 'We decided X because Y' is more useful than 'X happened' - Use category to classify: technical, decision, preference, identity, relationship - Use layer for context depth: 0=identity(stable), 1=architecture(monthly), 2=state(weekly), 3=session(daily) Before ending a conversation, call store_session_summary to preserve continuity.

Store up to 50 facts in one call — the same fields as store_fact per item. Made for carrying memory over: when you already remember things about the person from your own platform memory (ChatGPT memory, Claude memory, Gemini saved info), write them here one fact per item, in their words, with the date learned when known, so the memory follows them to every other tool. Each item is stored independently; the reply lists what landed and what did not.

Store a compressed summary of this session to preserve continuity for the next one. Call this before ending a conversation — it is the dream cycle. Stores a L3 session fact with key='session-latest' (or 'session-latest:claude-code' etc.), auto-archiving the previous version. The next session loads this via load_context(). Include: what was discussed/decided, what changed, what's next/unresolved. Collaborative takeaways: at natural session-end signals (user says 'okay I'm out', 'let's wrap', 'bye'), draft a one-line takeaway and confirm with the user before storing. Example: "I'll remember: we fixed the rebind bug and decided to skip transcript ingestion for MVP. Sound right?" Edit based on their response, then call store_session_summary. This makes memory consensual and catches mistakes immediately. Use session_id to scope by client: 'claude-code', 'claude-chat', 'cursor', 'openclaw'

Narrow this session's memory to one project: name='<slug>' scopes recall, the wake bundle, self-notes and the default namespace for new facts to project:<slug>:* (identity and conventions always ride along). Returns the narrowed bundle right away. name='' widens back to the whole account. Sticky for the session and across reconnects; bind_agent is unaffected. Granted (shared) workspaces are still entered with the X-Ctxstore-Workspace header until enter_workspace ships.

Check which context layers loaded successfully on this session start. Returns a structured status: which of embryo, conventions, agent identity, closing note, and opening note were found, plus a missing[] list. Call this after load_context() or resume_session() to verify your bootstrap was complete and diagnose what to fetch manually. Optional: pass profile='worker-dev' (or another platform:profile:<name> slug) to additionally check the REQUIRE keys named in that profile's recipe. Missing required facts surface in the same missing[] list.

Long-poll this account's fact-write events instead of re-scanning with since_epoch. Every store/supersede/delete/archive/restore/move publishes one event {key, id, namespace, layer, op, ts, seq}. One call waits up to max_wait_s (max 25) for the first matching event and returns up to 50, plus next_since. MCP has no streaming here, so loop: pass next_since back as `since` on the next call and nothing is missed (the server keeps the last ~10000 events per account). Omit since to start from now; since=0 replays what is retained; a bare integer N returns events with seq > N. EXAMPLES - Wait for hands to post results: watch_events(prefix='task:pr27-review:result:', max_wait_s=25) - Resume: watch_events(prefix='comms:<your-agent-id>:', since='<next_since>') Inside a workspace only events for keys the grant can read are returned. Long-lived clients can use GET /api/v1/events (SSE) instead.