A memory graph for AI that remembers why, not just what

What it does

AutoMem is a self-hosted memory service for AI assistants that stores typed relationships in a FalkorDB graph and vector embeddings in Qdrant. When you ask why you chose PostgreSQL, it surfaces the decision, the options you rejected, and the reasoning that tied them together—because it traverses relationships, not just keyword indexes.

The interesting bit

The project treats memory like biology, not bureaucracy: daily decay cycles, weekly REM-style connection discovery, and monthly clustering strengthen important links and let noise fade. It also performs multi-hop bridge discovery across its own graph to find the implicit logic connecting two distant memories.

Key highlights

• Dual storage: an 11-relationship-type graph in FalkorDB and 1024-dimension vectors in Qdrant; if vectors go down, the graph still answers in degraded mode.
• Benchmark claims of 87.00% on LongMemEval full and 84.74% on LoCoMo full, with 97.00% recall@5.
• Neuroscience-inspired consolidation cycles—decay, creative, cluster, forget—drawn from research including HippoRAG 2, A-MEM, MELODI, and ReadAgent.
• MCP bridge for Claude, Cursor, ChatGPT, Codex, and ElevenLabs agents, plus a direct REST API.
• Runs fully offline via Docker Compose or deploys to Railway from pre-built images.

Caveats

• The README contains several screenshot placeholders that currently render as broken-image icons because the assets are not yet in the repo.
• BEAM benchmark numbers are explicitly flagged as diagnostic and not apples-to-apples against published leaderboards.
• The authors themselves direct users needing SOC2, HIPAA, per-agent isolation, or a polished managed SaaS to alternatives like Mem0, Letta, or Zep.

Verdict

Worth testing if you are a solo dev or small team who wants ownable, relational AI memory that runs on your hardware. Look elsewhere if you need enterprise compliance, a managed dashboard, or strict per-agent memory isolation.