agentic-memory-stack-for-claude-code

Add a memory to mem0. Set infer=False to store as-is; True to LLM-extract facts. metadata dict can include {source, tier, kind, workspace, project, ...}.

Semantic search over mem0. Returns up to `limit` memories above `threshold`. Auto-reranks via the bge cross-encoder when limit>=5 (measured 2026-06-22 to IMPROVE relevance — blind Codex A/B 7/11; adds ~2s on this CPU box, fine for a DELIBERATE search, and NOT on the <500ms per-prompt bundle). Pass r…

PROACTIVELY pull the memory the task you are about to start depends on — the SAME admission-gated, brand-scoped, ranked context the per-prompt UserPromptSubmit hook USED to inject before that hook went dead in this runtime (VS Code extension / Agent SDK). Call it at the START of any substantive task…

List recent memories for user_id. Hard-capped server-side at 500. Prefer memory_search for content discovery; this is for inventory.

Delete a memory by its ID.

Update a memory's text content.

Promote a memory's trust tier. tier in {evidence, stable, insight, temporal}.

Demote a memory's trust tier (e.g., stable -> evidence when wrong). For full removal use memory_delete. Writes a tier-ledger entry. actor is always 'claude-autonomous' when called via MCP. reason is recommended for audit clarity.

Check mem0 server health: store, embedder, and canonical-key end to end.

Search past episodic memory (sessions) by keyword + optional date range + brand filter.

Return the last N episodes ordered by ended_at descending.

Fetch full detail for one episode by its integer id.

List goals with optional status/brand filters. status ∈ {open, blocked, advanced, completed, abandoned}. Use for 'show me all open goals for MyApp' / 'what's the goal landscape?'.

Recursive goal tree (adjacency-list hierarchy). If root_id is None, returns all top-level goals + descendants with depth field. Use for 'show me the goal hierarchy' / 'what's the structure of my plans?'.

Convenience: list open goals. Use for 'what am I currently working on?'.

Convenience: list blocked goals. Use for 'what's stuck right now?' / 'show me blockers'.

Full goal detail including linked_episode_count and parent reference. Use to investigate a specific goal.

Create a top-level or child goal directly (operator-direct, not auto-extracted from sessions). Use when the operator says 'add a goal: ship X by Y'.

Abandon a goal as no-longer-relevant. Distinct from 'completed' (which means goal was achieved). Use when scope changed, priority dropped, or goal was redundant with another. Requires a non-empty reason — this is a deliberate trash-can move; document why.

Mark a goal as completed — the work shipped / the goal was achieved. Distinct from 'abandon' (which means scope dropped or the goal became infeasible). Use this when a goal's deliverable is actually done so it stops bleeding into every session's surfaced context. Sets completed_at and stamps a goal-…

Exact read by memory_id. Returns text, metadata, tier, retrievable, source, timestamps.

W5 (ADOPT-2): why does memory `target_id` not surface for `query`?

Update a goal's priority (1=highest, 5=lowest). v0.17 F.3.2. Use when a goal's urgency changes relative to other open goals.

Explicitly link an episode to a goal. link_type ∈ {advanced_goal, blocked_goal, completed_goal, cited_goal}. Use when a session advanced a goal but auto-extraction from the Stop hook missed it. v0.17 F.3.2.

Merge source goal into target: move all episode links + mark source as 'duplicate'. Use when two goals turn out to be the same thing (duplicate created by auto-extraction). Source stays in DB for audit but is excluded from default listings. v0.17 F.3.2.

List current open frontier questions (Epistemic Reachability signal). Filter by brand. Default limit 7 = Miller's working-memory anchor.

FTS5 keyword search across open questions. status='all' to include resolved.

Mark a frontier question as resolved with a resolution summary.

Full detail on a frontier question including related goal title.