MemoryAtlas

A Flask REST API backed by two storage layers: FalkorDB holds memories as graph nodes connected by eleven authorable typed relationships plus three system-added edge types, while Qdrant holds a 1024-dimension embedding per memory. Recall blends semantic similarity, graph traversal, temporal alignment, tag overlap, and importance into a nine-component score, with multi-hop bridge discovery that surfaces the memory connecting two seed results rather than just the seeds themselves. Background enrichment and biologically-inspired consolidation cycles (daily decay, weekly creative linking, monthly clustering, optional forgetting) run continuously.

Storage
FalkorDB as the canonical graph record and source of truth, plus Qdrant for vectors. Recall degrades gracefully if Qdrant is unavailable; if FalkorDB is down the service returns 503, making the graph the hard dependency.
Retrieval
Hybrid nine-component scored recall across semantic similarity, graph traversal, temporal alignment, tag overlap, and importance, plus multi-hop bridge discovery across the typed relationship edges. Tags act as a hard pre-filter.
Self-host
Self-host: moderate
License
MIT
Pricing
The software is free and MIT-licensed; the only cost is infrastructure you run it on, which the project's own docs put at roughly $0.50–1/month on Railway after trial credits. No paid AutoMem tier exists — the README explicitly points enterprise, SOC2, and multi-tenant needs at Mem0, Letta, or Zep instead. · Free / OSS
GitHub stars
795
Last release
2026-07-07
Last commit
2026-07-18
First catalogued
2026-07-25

Strengths

  • True graph-plus-vector hybrid with eleven authorable relationship types and multi-hop bridge discovery — a genuinely differentiated retrieval mechanism, not flat similarity search
  • One installer wires up eight-plus clients, plus a remote HTTPS bridge for web-based assistants
  • Unusually self-skeptical benchmark posture: the README leads with the neutral third-party board and explicitly demotes its own self-judged numbers, naming the judge model and the roughly twelve-point swing risk
  • Documented, tunable consolidation cycles (decay, creative linking, clustering, forgetting) with explicit graceful-degradation behaviour per storage layer
  • Two clear deployment paths — one-click Railway or `make dev` Docker Compose

Watch out

  • Self-hosting requires three orchestrated services (Flask, FalkorDB, Qdrant) — there is no embedded or single-binary mode
  • The headline Agent Memory Benchmark figures (LoCoMo 85.1%, BEAM 10M 57.4%) were run and submitted by AutoMem's own team, and the submission PR to the neutral leaderboard remains open and unmerged as of 2026-07-25
  • The README's own limitations section is candid about real gaps: no SOC2 or HIPAA, no ACLs, no per-agent isolation for multi-agent swarms, tags acting as a hard pre-filter that can hide correct matches, and unreliable temporal conflict resolution
  • Explicitly pre-1.0
  • The MCP client is versioned and released separately from the core service, so there are two moving parts to keep in sync

Best for

  • Solo developers and small teams wanting one shared memory across several coding-agent tools via MCP
  • Users who specifically want typed-relationship reasoning rather than similarity search alone
  • Teams comfortable running Docker or Railway infrastructure who value an honestly-benchmarked self-hosted backend

How it integrates

Benchmark results

No sourced results yet.

Sources

Last verified 2026-07-25 · updated by discover-frameworks