Memory Engine MCP

Find atoms semantically similar to a given atom. Uses cosine similarity on embeddings. Args: atom_id: The reference atom limit: Max results threshold: Minimum similarity score (0-1) Returns: JSON string with similar atoms and scores

Get full atom details including all bonds (incoming and outgoing). Bumps access count.

Import memory data from a JSON export. Args: json_data: JSON string from export_all, or path to a .json file mode: "merge" (upsert by ID) or "replace" (wipe + insert) Returns: JSON string with import statistics

Import markdown file(s) into the memory engine. Args: filepath: Specific file to import, or None to import all markdown files Returns: JSON string with import statistics

Run the learning engine: detect contradictions, weak atoms, merge candidates, decay, and content gaps. Creates human questions for findings. Graph-gap/isolation review is handled by curator_run by default.

Create a typed bond between two atoms. Relations: is_a, part_of, depends_on, contradicts, refines, derived_from, detail_of, related_to

List atoms with optional filters. Useful for browsing the memory.

List explicit contradiction/supersession records. Args: atom_id: Optional atom id to filter old/new side of contradictions. limit: Max records to return.

Admin jobs pointer.

Answer a pending question: resolves it and optionally updates the linked atom status.

List pending human questions for review.

Create a SQLite backup (VACUUM INTO) with timestamp. Returns backup path and size.

Infer the 3-tier memory class (episodic/semantic/procedural) for a prospective atom.

Detect duplicate durable atoms by normalized title (report only, or auto-merge with auto_apply).

Purge expired session_msg/session_digest atoms (TTL passed). Returns purged count.

Cognitive health metrics for the memory graph.

New atom supersedes old one.

Conservative curation pass: compact, bonds, isolated classification, promotions, merges. dry_run default true.

Decay cycle: reduce weight of atoms not accessed within interval_days by factor (default 0.95). Returns decayed count.

Delete an atom: soft (status='deleted', default) or hard (remove row + bonds + versions). Hard delete is irreversible.

Known errors for a task.

List recorded error memories.

Log mistake + correction.

Export all atoms + bonds to a JSON file. Returns file path and counts.

Export a single atom (with bonds) to JSON. Returns the data.

Find atoms similar to a reference atom. Semantic via embeddings (cosine) with FTS fallback.

Generate digest atom for a session.

Full atom by id.

What depends on an atom.

Import atoms from a JSON export file (merge mode: skip existing IDs). Returns import counts.

Import atoms from a markdown file: each heading (##/###) becomes an atom, body is the content.

Run learning checks (full Python parity): contradictions, weak atoms, merge candidates, decay critical, gaps (body vuoti), graph gaps (opt-in) + stale_fact/needs_link. Creates deduped human_questions; answers apply side effects.

Bond two atoms.

List atoms by domain/type.

List available backup files with sizes and dates.

List contradiction records (superseded atoms).

Merge duplicate atoms into a primary: move bonds, mark duplicates merged (soft). Returns merge report.

Search stable preferences.

Smart recall of past memories. Query short and specific. semantic=true merges embedding search.

Recall messages from a specific OpenClaw session (search within session/ domain atoms).

(Re)generate embeddings for active durable atoms missing one. Batch with max_atoms; returns indexed/failed counts.

Save a durable memory (fact, decision, procedure, preference).

Restore from a backup file. DANGEROUS: overwrites current DB. Requires confirm=true.

Add a binding rule.

Permanently delete a binding rule by id.

Get active binding rules.

List all binding rules.

Deactivate a binding rule by id.

Activate or deactivate a binding rule.

BFS graph traversal.

Semantic search over atom embeddings (Ollama). Falls back to FTS with a note when embeddings/Ollama are unavailable.

Digest of a session.

Memory statistics.

Suggest bonds for one atom (domain cluster, keyword overlap, id patterns). auto_apply=false by default.

Bulk bond suggestion across active atoms (fast strategies only). auto_apply=false by default.

Hierarchical 3-level memory summary: L0 global counts, L1 per-domain top atoms, L2 pointer to detail.

Remove a bond between two atoms.

Bulk update the domain of one or more atoms.

Memory Engine TS version.

Task context pack with token budget.

Merge two atoms. Secondary is marked as 'merged', all bonds move to primary. Primary gets a confidence boost.

Search preference atoms by structured metadata. Preferences are atoms with type='preference' and metadata like: {"category": "tooling", "condition": "when compiling PS", "rule": "dotnet build net48"} Args: category: Filter by category (e.g. 'tooling', 'naming', 'format', 'policy') query: Free-text search in title, body, and meta fields scope: Filter by scope (e.g. 'personal', 'project') limit: Max results Returns: JSON string with matching preferences

Smart recall: search memory with multi-factor ranking. Combines FTS relevance, semantic similarity, confidence, recency, and weight. Args: query: Natural language query limit: Max results (default 5) min_weight: Filter out low-weight atoms domain: Filter by domain semantic: If True (default), also use semantic search via Ollama memory_tier: Optional filter: episodic, semantic, procedural include_superseded: Include historical superseded atoms (default False) Returns: JSON string with ranked results

Recall messages from a specific OpenClaw session. Searches within session_msg atoms for the given session. Args: session_id: Session ID (filename without .jsonl) query: Search query to filter messages limit: Max results

Check the status of the background embedding indexer. The indexer runs automatically at startup and processes all atoms missing embeddings. Args: force: If True, delete all existing embeddings and restart from scratch Returns: JSON string with current indexing progress

Create or update a memory atom. Args: title: Short, clear title for this memory body: Full content (can be markdown) type: One of: fact, decision, event, preference, log, procedure, note domain: Categorization (e.g. 'infrastructure', 'personal', 'project:xxx') confidence: 0.0 (hypothesis) to 1.0 (verified) tags: List of tags for categorization ttl_days: Optional TTL in days (None = permanent) memory_tier: Optional 3-tier class: episodic, semantic, procedural Returns: JSON string with created atom info

Restore database from a backup file. WARNING: This overwrites the current database. A safety backup of the current state is created automatically before restore. Args: backup_path: Path to the backup file to restore from Returns: JSON string with restore result

Traverse the knowledge graph starting from an atom. Args: atom_id: Starting atom depth: How many hops to traverse (default 2) relation: Filter by specific relation type Returns: JSON with nodes and edges discovered.

Semantic search using Ollama embeddings (nomic-embed-text). Finds atoms by meaning, not just keyword match. Args: query: Natural language query limit: Max results (default 5) domain: Filter by domain min_weight: Filter by minimum weight Returns: JSON string with ranked semantic results

Get a summary of a session: message count, time range, topics. Useful for quickly understanding what was discussed. Args: session_id: Session ID (filename without .jsonl)

Get memory engine statistics: atom counts, bonds, pending questions, breakdown by domain and type, average weight, low-confidence count.

Suggest bonds for an atom using rules + semantic similarity. Strategies: domain cluster, keyword overlap, pattern detection, semantic. Args: atom_id: The atom to find bonds for auto_apply: If True, create the bonds directly Returns: JSON string with bond suggestions

Scan active atoms and suggest/create bonds. Useful for initial graph population or periodic maintenance. Args: auto_apply: If True, create bonds directly limit_per_atom: Max suggestions per atom max_atoms: Max atoms to scan (default 30) Returns: JSON string with summary statistics and top suggestions

Remove a bond between two atoms.

Update the domain of one or more atoms (bulk-friendly). Useful to consolidate fragmented domains (e.g. 'projects/x' vs 'project:x'). Does NOT touch embeddings or FTS (domain is not part of the indexed text). Args: atom_ids: list of atom ids to move new_domain: target domain include_archived: also update non-active atoms (default False) Returns JSON: {updated, not_found, skipped_archived, new_domain}

Get the Memory Engine server version.

Build a task-oriented context pack before doing work. Combines smart recall, graph neighbors, and key procedures/decisions into a compact working set so the agent starts with the right operational memory.