mem01
Long-term memory for AI agents
A self-hosted memory service that tracks changing facts, filters stale beliefs, and retrieves useful context within a token budget.
Local v0.3 benchmark on 1,540 questions.
Remember a preference.
Interactive illustration, no service calls.
How I built it : mem01
The problem. Agent memory accumulates stale facts and contradictions. Retrieving more context does not help when that context contains outdated information.
My approach. Store memory as beliefs with an explicit lifecycle: add, update, supersede, invalidate, and merge. Retrieval filters conflicts and packs the results into a token budget, with zero LLM calls on the read path. The self-hosted service uses FastAPI, PostgreSQL, and pgvector, packaged with Docker Compose.
The result. The v0.3 local benchmark scored 80.5% on LoCoMo (1,240 of 1,540 non-adversarial questions), using gpt-4o-mini and mem0's open methodology. Mean packed context was 1,252 tokens. Recall latency was 442ms p50 and 913ms p95.
The limits. This is a single local benchmark run, with roughly two percentage points of judge noise. The 82% ship gate was not met, so the evaluated v0.3 remained local. It did not match mem0's published 92.5% score.
Python · FastAPI · PostgreSQL · pgvector · Docker
