r3 MCP server

Read from the package source without running it. The installed server may list more.

• add_memory

  Store a new memory with automatic deduplication and indexing. Use for persisting facts, preferences, or conversation context. Checks for duplicates by default (85% similarity threshold). Returns immediately; background processing handles indexing. Prefer over update_memory for new content. Returns: confirmation text. Side effects: creates memory record, updates search index, may skip if duplicate

• cache_stats

  View cache performance metrics and health status. Use for monitoring and debugging. Shows memory count, access patterns, and hit rates. Read-only. Returns: summary text with cached memory count. No side effects.

• deduplicate_memories

  Detect and optionally remove duplicate memories using content similarity. Run with dry_run=true first to preview. Compares all memories pairwise using Jaccard similarity. Groups duplicates with a primary (oldest) and candidates for removal. Returns: summary with duplicate groups. Side effects (when dry_run=false): deletes duplicate memories, invalidates cache.

• delete_memorywrite action

  Permanently remove a memory by ID. Irreversible operation. Removes from storage, cache, and search index. Use deduplicate_memories with dry_run first to preview bulk deletions. Returns: confirmation text. Side effects: deletes memory record, removes from all indexes, invalidates cache.

• extract_entities

  Extract named entities and relationships from text using NLP. Identifies people, organizations, technologies, and projects. Also extracts relationships (WORKS_FOR, USES, etc.) and keywords. Requires enhanced intelligence mode. Read-only, stateless. Returns: {people[], organizations[], technologies[], projects[], relationships[], keywords[]}.

• find_connections

  Discover relationship paths between entities in the knowledge graph. Uses BFS traversal to find how entities connect. Useful for answering 'how is X related to Y?' questions. Requires enhanced mode. Read-only. Returns: {from, to, max_depth, paths_found, paths[]} where each path is array of {from, to, type} edges.

• get_all_memories

  List all memories for a user with pagination. Use for browsing or bulk operations. For content search, use search_memory instead. Cache-first by default. Large result sets are automatically truncated. Read-only. Returns: {total, limit, offset, returned, hasMore, source, memories[]}.

• get_knowledge_graph

  Build a knowledge graph from stored memories. Returns entities as nodes and relationships as edges. Use for visualizing connections between concepts. Requires enhanced intelligence mode with prior entity extraction. Read-only. Returns: {nodes[], edges[]} where nodes have {id, type, name, memories[]} and edges have {from, to, type, confidence}.

• get_memory

  Retrieve a single memory by its unique ID. Use when you have a specific memory_id from prior search/list results. Returns null if not found. Prefer search_memory for content-based lookup. Read-only operation. Returns: Memory object {id, content, user_id, metadata} or null.

• import_memories

  Bulk import memories from external sources. Supports Mem0 API export or local JSON files. Processes in batches with duplicate detection. Use for migration or backup restoration. Returns: summary with imported/skipped/failed counts. Side effects: creates multiple memory records, updates search index.

• optimize_cache

  Reorganize cache for optimal hit rates. Promotes frequently accessed memories to L1 (24h TTL), demotes cold data to L2 (7d TTL). Use periodically for large memory stores. May temporarily increase latency during optimization. Returns: summary of cached memories. Side effects: modifies cache TTLs, may evict old entries.

• search_memory

  Find memories matching a natural language query using hybrid semantic + keyword search. Primary retrieval tool for content-based lookup. Uses vector similarity (enhanced mode) or keyword matching (basic mode). Cache-first by default for speed. Returns ranked results with relevance scores. Read-only. Returns: array of Memory objects or 'No memories found' text.

• sync_status

  Check background job queue and sync status. Shows pending async operations from add_memory calls. Use to verify all writes completed. Read-only. Returns: count of pending operations or 'All operations complete'. No side effects.

• update_memorywrite action

  Modify an existing memory's content or metadata. Use for corrections or adding context to existing memories. Fails if memory_id not found. Prefer add_memory for new content. Invalidates search cache. Returns: updated Memory object. Side effects: modifies memory record, invalidates cached search results.

• Code scan60 source files scanned25/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitystatic API keys via environment variables6/15
• Maintenancelast push 1 days ago15/15
• Maintainer identityregistry namespace matches repository owner; GitHub account older than a year8/10

r3: common questions

Is r3 MCP server safe?

Mostly: it is graded B (83/100). Read the r3 safety report

How do I install r3?

It runs on your machine. Copy the Claude Code, Claude Desktop or Cursor config from the install section.

Does r3 need an API key?

Yes. The registry entry asks for MEM0_API_KEY.

Is r3 maintained?

The last commit was 2 days ago (2026-09-18). The latest release is v1.3.2.

What can I use instead of r3?

Servers from other publishers that do the same job: Redis MCP Server, KEPTA — Agent Memory MCP server and Memory Vault MCP server. Compare all r3 alternatives.