MemoryAtlas

Memory in plain SQL — no vector DB, fully inspectable, portable. LoCoMo benchmark: 81.95% accuracy at ~1,294 tokens/query (self-reported). Paper: arxiv.org/abs/2603.19935.

Storage
Any SQLite / Postgres / MySQL; BYODB option supported for self-hosted databases
Retrieval
Recalls memories via plain SQL queries over relational tables — no vector database required; works with any LLM + datastore combination.
Self-host
Self-host: trivial
License
Apache-2.0
Pricing
Free OSS; managed cloud Free/Paid · Free + paid
GitHub stars
15,486
Last release
2026-05-28
Last commit
2026-06-15
First catalogued
2026-06-28

Strengths

  • SQL-queryable / fully debuggable memory
  • ~1,294 tokens/query on LoCoMo (self-reported) — ~5% of full-context cost
  • One-line SDK integration: .llm.register(client)
  • BYODB: bring your own SQLite / Postgres / MySQL
  • Captures agent actions (tool calls, decisions) not just conversation text
  • OpenClaw plugin available

Watch out

  • Simpler retrieval than graph/hybrid systems — may miss relational or temporal queries
  • Self-reported benchmark (81.95% LoCoMo); backbone LLM not stated in source

Best for

  • Cost-sensitive production: skip the vector DB and run on the SQL infra you already have
  • Inspectable, debuggable memory you can query directly
  • Agents where what the agent *did* (tool calls, decisions) matters as much as what the user *said*

Benchmark results

BenchmarkValueBackboneTrustSource
locomo81.95 accuracySelf-reportedMemori (MemoriLabs)

Sources

Last verified 2026-06-29 · updated by seo-article-writer