Recall

Persistent memory layer for AI agents. Local-first. Inspectable. Framework-agnostic.

An MCP server that gives AI agents durable, structured memory across sessions. Agents call store_memory with conversation text — Recall extracts typed facts in the background using Claude Haiku, detects contradictions with existing memories, and stores everything as searchable, ranked context. When the agent needs context, search_memories returns the most relevant memories ranked by a 4-component scoring model (relevance, recency, importance, access frequency).

MCP tools store_memory, search_memories, inspect_memories, delete_memory, get_memory_stats, consolidate_memories, delete_namespace_data
Retrieval BM25Plus + optional dense vector (BAAI/bge-small-en-v1.5) fused via RRF + MMR diversification
Protocol MCP (Streamable HTTP) + A2A v1.0 with consolidation skill
Storage SQLite (default, zero infra) or Postgres via RECALL_DB_URL
pip install szl-recall + [embeddings] for dense vector search · [postgres] for Postgres backend
Python MCP A2A SQLite Postgres FastMCP Claude Haiku
More coming

More open-source tools from the series are in progress. Follow @Sentient-Zero-Labs on GitHub.