Memory system architecture used in Mastra's Observational Memory with two background agents that compress and garbage collect conversation history achieving 5-40x compression.
Loading more......
The Observer-Reflector Architecture is a two-agent system used in Mastra's Observational Memory that watches conversations and maintains a compressed text-only observation log.
When messages hit 30k tokens (configurable threshold), a separate "observer agent" compresses messages into new observations that are appended to the observation block. The compression is typically 5–40×.
When observations hit 40k tokens (also configurable), a separate "reflector agent" garbage collects observations that don't matter, removing outdated or irrelevant information.
The context window is broken into two blocks:
Observational Memory achieves 94.87% on LongMemEval with GPT-5-mini, the highest score ever recorded on this benchmark.
Open-source as part of Mastra framework.