Agentlas hits 96.4% on LongMemEval without calling a model for retrieval
minglingj · reddit · 2026-07-25
A new project called Agentlas claims 96.4% accuracy on LongMemEval without calling an LLM during retrieval, and 91.5% on LoCoMo.
Its memory design stays local in a single SQLite file rather than a separate memory server. The system splits memory into four ownership zones: user identity, project context, individual agent learning records, and team-shared facts. Instead of asking an LLM every time, it combines a lightweight on-device embedding model with keyword search to filter irrelevant results and reduce hallucinations. The latest update also adds independent per-member identifiers, plus a self-evolving mechanism that can propose improvements after repeated failures and let users approve or rollback them with one click.
More from coding & agent
- Shopify design lead on AI-era workflow: designers are becoming full-stack editors — KlausCodes · 2026-10-26
- Ultrametric open-sources Open Startup Repo to let AI agents run startup operations — garrytan · 2026-10-26
- MCP fault injection study: agent task scores drop 60% under realistic tool failures, scale doesn't help — dair_ai · 2026-10-26
- General agents burn ~140M tokens/month on LinkedIn outreach; scripts do it 60x cheaper — ChrisHarpon2 · 2026-10-26
- OpenAI upgrades Decisions API: 400 questions per call, native voice and image support — OpenAIDevs · 2026-10-26
- NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26