[Idea] Integrating Long-Term Temporal Memory (Smriti) for stateful computer-agent sessions
#25
by Spy9191 - opened
Hey team!
The computer-agent space is incredible, but I noticed that agent sessions often lose context or hallucinate past facts if the container restarts or the session runs too long.
Has the team considered separating the episodic execution memory from a durable semantic memory layer? We built an open-source bi-temporal database called Smriti that natively solves this. It extracts SVO events from the chat and gives them explicit valid_from and valid_to bounds, automatically expiring old facts without slow LLM calls.
It could be a really powerful plug-and-play backend for smolagents if you want to support true stateful memory. Happy to help put together a quick PR or share our PostgreSQL schemas if interested!