Announcing Observational Memory

A human-inspired memory system that scores ~95% on LongMemEval — with a completely stable context window.

Observational memory is text-based (no vector/graph DB needed), SoTA on benchmarks like LongMemEval, and compatible with Anthropic/OpenAI/etc prompt caching.

Observational memory works like human perception. On a busy street, your brain processes millions of pixels but distills them down to one or two observations. In the context of a coding agent, it compresses a user session down to key observations, keeping the context window stable while retaining what matters.