ClaudioDrews/memory-os

Building memory for agents is one of those problems where every team reinvents a slightly different broken wheel. This open-source project ships a complete 7-layer memory OS — persistent vector storage with Qdrant, structured fact storage, a fabric-style recall system, an auto-curated wiki layer, and surgical context injection that puts only the relevant memory slice into the prompt window. The local-first, any-provider stance is the real differentiator: no API dependency, no memory vendor lock-in, no cloud bill that scales with context size. Most competing memory libraries paper over the hard parts — retrieval ranking, context budget management, fact deduplication. This one has clear architectural layers for each concern, which means it is debuggable when it misbehaves in production. Reservation: it is built for the Hermes Agent, so adaptation to other agent frameworks will require reading the internals and doing some wiring yourself. That said, the architecture is worth studying even if you end up porting only the ideas. -> Best for: AI engineer building a production agent that needs durable memory without a managed memory API.