Agent Memory Service — temporal knowledge graph with structured episodic recall, hybrid retrieval, and progressive disclosure. ~600 lines. One database.
FalkorDB handles graph, vector, and fulltext search in a single database. Graphiti extracts temporal knowledge automatically.
Cosine + BM25 + graph traversal
Bi-temporal entities & edges
Precision reranking pass
Graphiti + FalkorDB. Thin FastAPI wrapper. ~600 lines of custom code.
Cosine similarity, BM25 keyword matching, and graph traversal (BFS) combined for maximum retrieval accuracy. Cross-encoder reranking for precision.
Bi-temporal edges track when facts become valid and when they expire. Your AI understands not just what happened, but when — and whether it's still true.
Memory types ranked by salience: Facts (5), Decisions (8), Failures (9), Reflections (9). Higher salience surfaces first — critical lessons before routine context.
From raw interaction to structured knowledge graph. Graphiti runs the full extraction pipeline: entities, relationships, temporal edges, deduplication.
Agent sends episode via /v1/ingest
LLM extracts entities & relationships
valid_at / invalid_at timestamps
Cosine + BM25 + graph + reranking
"Intelligence without memory is just computation."