// the find
verygoodplugins/automem
AutoMem is a graph-vector memory service that gives AI assistants durable, relational memory:
AutoMem is a self-hosted memory service for AI assistants that stores memories as typed graph edges in FalkorDB and semantic embeddings in Qdrant, then queries both layers together at recall time. The multi-hop graph traversal — finding the bridge node that connects two semantically related memories — is the actual differentiator over plain vector search. Aimed at solo developers and small teams who want cross-client persistent memory and are comfortable running two databases.
The hybrid recall is the real thing: 11 typed relationships (PREFERS_OVER, EXEMPLIFIES, CONTRADICTS, etc.) let you traverse meaning that cosine similarity misses entirely. The benchmark methodology is honest — experiment log, postmortems for regressions, and a publication bundle with reproducibility commands, not just a headline number on the README. Graceful degradation is well-designed: Qdrant down triggers degraded-mode graph recall, FalkorDB down returns 503 and makes the failure mode explicit. The MCP bridge as a separate npm package with per-client setup templates is a practical integration story that actually works with Claude Desktop and Cursor without any glue code.
Two databases to operate (FalkorDB + Qdrant) is real overhead — FalkorDB is Redis-wire-compatible but has a small user base and thin ecosystem, so you're betting on a niche dependency for your canonical data store. The module naming is a red flag: nearly every file is prefixed runtime_* (runtime_helpers.py, runtime_bindings.py, runtime_recall_helpers.py, runtime_routes.py) which is either heavy over-engineering for the codebase size or an artifact of a large refactor that was never cleaned up. No per-user or per-agent memory isolation — all memories share one graph, so you cannot run this for multiple users without building that layer yourself, which the README admits but undersells as a limitation. The screenshots section in the README explicitly acknowledges there are no screenshots yet, which undercuts confidence in the graph visualization claim.