Delete multiple memories matching filter criteria. PERMANENT deletion.

OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'run_id' - filter by session/run identifier OPTIONAL: 'metadata_type' - delete memories with this metadata type (e.g., "test", "note") OPTIONAL: 'older_than_days' - delete memories older than this many days OPTIONAL: 'dry_run' - if True, returns what would be deleted without actually deleting (default: True)

⚠️ WARNING: Set dry_run=False to actually delete. This cannot be undone.

Examples:

• Preview test cleanup: {"metadata_type": "test", "dry_run": true}

• Delete old memories: {"older_than_days": 90, "dry_run": false}

• Delete by type: {"metadata_type": "note", "dry_run": false}

Returns: List of deleted (or would-be-deleted) memory IDs.

Delete all memories for a specific scope. PERMANENT deletion.

OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'run_id' - delete all memories for this session/run

⚠️ WARNING: This permanently removes ALL memories matching the scope. Cannot be undone.

Examples:

• Delete all user memories: {"user_id": "john"}

• Delete session memories: {"run_id": "session_123"}

Returns: Confirmation of deletion with count.

Delete memory by its unique ID. PERMANENT deletion.

REQUIRED: 'memory_id' - the UUID of the memory to delete

⚠️ WARNING: This permanently removes the memory and cannot be undone.

Export memories to JSON or Markdown format for backup or migration.

OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'format' - output format: "json" (default) or "markdown" OPTIONAL: 'include_metadata' - include metadata in export (default: True) OPTIONAL: 'output_path' - custom file path (default: .mem0/exports/memories_.) OPTIONAL: 'return_content' - if True, returns content instead of writing to file (default: False)

Returns: File path where export was saved, or content if return_content=True.

Use for: Backup, migration between backends, documentation, sharing.

Get specific memory by its unique ID.

REQUIRED: 'memory_id' - the UUID of the memory to retrieve

Returns: Complete memory details including content, metadata, timestamps.

Get change history for a specific memory by ID.

REQUIRED: 'memory_id' - the UUID of the memory to get history for

Returns: Chronological list of all changes made to the memory.

Get statistics about stored memories.

OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user)

Returns: Memory counts by metadata type, date ranges, and storage summary.

Use for: Understanding memory usage, identifying cleanup opportunities, monitoring growth.

Get recently added or updated memories for session continuity.

OPTIONAL: 'days' - how many days back to search (default: 7) OPTIONAL: 'limit' - max memories to return (default: 10) OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'run_id' - filter by session/run identifier

Returns: Most recently added/updated memories sorted by newest first

Check the health and connectivity of the memory backend.

Returns: Backend type, connection status, and basic diagnostics.

Use for: Troubleshooting connection issues, verifying configuration.

Import memories from exported JSON file.

REQUIRED: 'file_path' - path to the JSON export file OPTIONAL: 'user_id' OR 'agent_id' - override the user/agent from the export file OPTIONAL: 'skip_existing' - if True, skip memories that already exist (default: True) OPTIONAL: 'infer' - use LLM processing during import (default: False for speed)

Imports memories from a JSON file created by export_memories tool. Useful for backup restoration, migration between backends, or sharing memory sets.

Examples:

• Import from backup: {"file_path": ".mem0/exports/memories_john_20250113.json"}

• Import to different user: {"file_path": "export.json", "user_id": "jane"}

• Force reimport: {"file_path": "export.json", "skip_existing": false}

Returns: Import summary with success/failure counts.

Ingest documents into memory as a knowledge base using LlamaIndex.

Supports: PDF, DOCX, TXT, MD, HTML, and more. Automatically chunks and stores content.

REQUIRED: 'path' - file path or directory path to ingest OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'run_id' - session/run identifier for partitioning OPTIONAL: 'recursive' - recursively process subdirectories (default: False) OPTIONAL: 'chunk_size' - max characters per chunk (default: 1024, min: 100, max: 8192) OPTIONAL: 'chunk_overlap' - overlap between chunks (default: 200, min: 0, max: chunk_size-1) OPTIONAL: 'file_metadata' - additional metadata to attach to all chunks OPTIONAL: 'infer' - use LLM for fact extraction (default: False for speed)

Examples:

• Ingest single file: {"path": "/path/to/document.pdf"}

• Ingest directory: {"path": "/path/to/docs", "recursive": true}

• With metadata: {"path": "manual.pdf", "file_metadata": {"type": "documentation", "version": "2.0"}}

• Custom chunking: {"path": "large.pdf", "chunk_size": 2048, "chunk_overlap": 400}

Returns: Summary of ingested documents with chunk counts and memory IDs.

List all memories for a user or agent with pagination.

OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'run_id' - filter by session/run identifier OPTIONAL: 'page', 'page_size' - pagination controls (default: page_size=25)

Returns: All memories belonging to the specified user/agent, sorted by creation date.

Reset the entire memory store. DESTRUCTIVE operation.

⚠️ CRITICAL WARNING: This deletes ALL memories in the backend, regardless of user/agent/app. Only use this for testing or when you need to completely clear the memory store.

This operation cannot be undone.

Returns: Confirmation of reset.

Search memories by metadata fields without semantic search.

OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'type' - filter by metadata type (e.g., "decision", "preference", "task_completion") OPTIONAL: 'priority' - filter by priority (e.g., "high", "medium", "low") OPTIONAL: 'status' - filter by status (e.g., "complete", "in_progress", "blocked") OPTIONAL: 'custom_filters' - dict of additional metadata key-value pairs to match OPTIONAL: 'page' - page number for pagination (default: 1) OPTIONAL: 'page_size' - results per page (default: 50, max: 100)

Returns: Memories matching all specified filters with pagination.

Examples:

• Find decisions: {"type": "decision"}

• Find high priority: {"priority": "high"}

• Find blocked items: {"status": "blocked"}

• Custom filter: {"custom_filters": {"category": "api"}}

• Paginated: {"type": "documentation", "page": 2, "page_size": 20}

Note: Uses native Qdrant filtering when available, falls back to post-filtering for other backends.

Search memories with semantic similarity and relevance filtering.

REQUIRED: 'query' - what to search for (natural language) OPTIONAL: 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) OPTIONAL: 'run_id' - filter by session/run identifier OPTIONAL: 'limit' - max results to return (default: 10, also sets page_size if not specified) OPTIONAL: 'page' - page number for pagination (default: 1) OPTIONAL: 'page_size' - results per page (default: uses limit value)

Returns: Relevant memories filtered by maximum score threshold (configurable via MEM0_MAX_RELEVANCE_SCORE, default: 1000)

Score interpretation (FAISS uses L2 distance - lower is more similar):

• 0-100: Very high similarity (near-exact matches)

• 100-400: High similarity (closely related)

• 400-600: Moderate similarity (somewhat related)

• 600-1000: Low similarity (loosely related)

• 1000+: Very low similarity (likely unrelated)

Note: Default threshold of 1000 keeps reasonably similar results. Lower to 400-600 for stricter filtering.

Examples:

• Search user memories: {"query": "React patterns", "user_id": "john"}

• Auto-detect user: {"query": "my project status"}

• Limit results: {"query": "API endpoints", "limit": 3}

• Session-scoped: {"query": "current task", "run_id": "session_123"}

• Pagination: {"query": "all tasks", "page": 2, "page_size": 5}

Use for: Finding relevant code, recalling user preferences, retrieving project context.

Setup the memory-first steering file for Kiro.

Creates a steering file at .kiro/steering/memory-first.md that instructs the AI to always check memory before starting tasks and store important outcomes.

OPTIONAL: 'workspace_path' - path to workspace root (default: current directory)

Call this once after adding the MCP server to enable memory-first workflows.

Returns: Success message with file path created.

Store memory content with metadata support.

REQUIRED: Either 'user_id' OR 'agent_id' (if neither provided, auto-detects current user) REQUIRED: 'content' - the information to remember

OPTIONAL: 'metadata' - structured data about the memory (JSON object) OPTIONAL: 'infer' - if True (default), uses LLM to extract facts and dedupe. Set False for faster raw storage. OPTIONAL: 'run_id' - session/run identifier for temporary context OPTIONAL: 'custom_instructions' - per-call instructions to control what gets stored OPTIONAL: 'verbose' - if True, return full details. If False (default), return compact response.

Examples:

• Store personal info: {"content": "User prefers React over Vue", "user_id": "john"}

• Store with metadata: {"content": "API endpoint changed", "metadata": {"type": "technical", "priority": "high"}}

• Fast storage (no LLM): {"content": "Quick note", "infer": false}

• Session-scoped: {"content": "Current task context", "run_id": "session_123"}

Use for: Storing code patterns, user preferences, project details, technical knowledge.

Note: With infer=True (default), mem0 uses an LLM to extract key facts and intelligently update existing memories. This is slower (~5-15s) but smarter. Use infer=False for faster raw storage (~1-2s) without deduplication.

Config: Set MEM0_INFER_DEFAULT=false in env to disable infer by default. Config: Set MEM0_VERBOSE=true in env to enable verbose responses by default.

Returns: Success message with memory ID (compact) or full details (verbose).

Update an existing memory's content and/or metadata directly.

REQUIRED: 'memory_id' - the UUID of the memory to update REQUIRED: 'content' - the new content for the memory

OPTIONAL: 'metadata' - new metadata to replace existing (JSON object)

Unlike store_memory with infer=true, this directly updates the specified memory without LLM processing or deduplication. Preserves memory history.

Examples:

• Update content: {"memory_id": "abc-123", "content": "Updated preference: Vue over React"}

• Update with metadata: {"memory_id": "abc-123", "content": "New content", "metadata": {"priority": "high"}}

Returns: Updated memory details.