Persistent memory for AI agents — REST API, MCP, OAuth, CLI, dashboard. Semantic search, knowledge graph, automatic consolidation. One self-hosted service, every transport.
Interactive 3D knowledge graph — every memory a glowing node, every relationship a curved edge.
MINOR release: 19 merged pull requests, most of them from outside the maintainer. Harvested memories now record where they came from, and three storage bugs that only surface at real sizes are closed. ~3,360 tests.
An insight now records how it was produced: a harvest:method:llm or harvest:method:heuristic tag, plus the model, the pipeline version and the session id in metadata, on the store path and the evolve path alike. verify_session_coverage builds on it — before you delete a source transcript, it re-harvests the file in memory, compares each insight against what is stored, and answers safe_to_delete with the ones it could not find. (#1243, #1252)
Tag-filtered retrieval on Cloudflare borrowed the sqlite-vec KNN ceiling for topK, so any request above 16 results hit Vectorize's metadata cap and raised a 4xx instead of returning. Hybrid pull sync compared counts rather than diffing hashes, so two divergent stores of equal size were reported in sync. And get_by_hash still returned soft-deleted rows on Milvus and Cloudflare, so a re-store saw a tombstone as a live duplicate. (#1259, #1255, #1262)
The NLI cascade reuses the harvest provider chain, skips the request entirely when no provider is configured, and emits one bounded warning per run instead of degrading to the heuristic without a word. Configuration got the same treatment: an unrecognized path variable such as MCP_MEMORY_DB_PATH was ignored in silence while the database landed on the default path, and now warns once, naming the variable it saw and the one it wanted. (#1265, #1253)
Claude.ai connects to your memory server through Streamable HTTP with OAuth 2.1 authentication.
Battle-tested with comprehensive testing and optimized for performance.
Talks and coverage from the community.
Heavybit's Humans in the Loop deep dive talks to maintainer Heinrich Krupp about agent amnesia and why persistent memory is the missing infrastructure layer for agentic systems.
Maintainer Heinrich Krupp presents mcp-memory-service at the AI Tinkerers Zürich meetup — persistent memory architecture, semantic search, and multi-agent memory sharing. Watch the full video →
Install from PyPI and connect your AI agents to persistent memory.
I help teams run self-hosted memory in production — OAuth, hybrid sync, multi-client rollout. Free 30-minute intro call.
Get in touch