MemoryAtlas

Fastest setup + largest integration ecosystem; dual session/user scope.

Storage
Cloud, or OSS lib + your vector store + LLM
Retrieval
LLM-extracted memories stored in a vector store (optional graph tier) and recalled by semantic search.
Self-host
Self-host: moderate
License
Apache-2.0
Pricing
Free tier; paid from ~$19/mo; graph ~$249/mo · Freemium
GitHub stars
59,620
Last release
2026-06-27
Last commit
2026-06-27
First catalogued
2026-06-28

Strengths

  • ~30s to wire up
  • 21+ frameworks
  • Generous free tier

Watch out

  • Self-reported benchmarks far exceed neutral ones; graph tier is pricey

Best for

  • Fastest drop-in memory with the biggest integration ecosystem
  • Token-cost-sensitive production agents (single-pass extraction, sub-7k tokens/call)
  • AWS Agent SDK users and SOC 2 / HIPAA workloads

Benchmark results

BenchmarkValueBackboneTrustSource
longmemeval94.4 accuracySelf-reportedMem0
longmemeval49 accuracyGPT-4oIndependentZep (competitor harness)
locomo92.5 accuracySelf-reportedMem0
beam-1m64.1 accuracySelf-reportedMem0
beam-10m48.6 accuracySelf-reportedMem0
locomo66.88 accuracyIndependentHindsight/Vectorize (competitor re-run)
longmemeval62.6 accuracygpt-4o-miniIndependentLiCoMemory (Huang et al., HKUST et al.) — competitor re-run
locomo54.68 accuracygpt-4o-miniIndependentLiCoMemory (Huang et al., HKUST et al.) — competitor re-run

Sources

Last verified 2026-06-28 · updated by refresh-framework-cards