memory-os

Why we included this project

Anyone running Hermes Agent in real projects knows the exact problem this project was built to solve: the agent resets every session and you have to re-teach it your stack, your decisions, and your reasoning. Memory OS sits underneath the agent as a local persistent memory layer. It keeps a vector store for semantic recall, structured facts with trust scores that improve as you work, a wiki it curates itself, and full-text search across past conversations. On each turn it injects only the context that is relevant right now, so token spend stays tight. It runs entirely on your own hardware and works with any LLM provider, so you get durable memory without a cloud subscription or vendor lock-in. A one-command installer plus bundled smoke and ingestion tests let a small team validate it on modest hardware before committing. If you have been manually pasting context into every prompt, this is a concrete way to make the agent behave like a colleague who was present for every prior session.