From an engineering perspective, what I find compelling here is not “no embeddings”, but the decision to treat memory as a first-class, inspectable system rather than a retrieval trick.
Most agent memory stacks today collapse everything into embeddings and hope similarity search is enough. That works for recall, but breaks down quickly when you need traceability, temporal reasoning, or explanation of why something was remembered.
The layered design here (raw resources → extracted memory items → categorized memory files) feels much closer to how we design real systems: separation of concerns, clear abstraction boundaries, and the ability to reason about state changes over time.
Storing memories in human-readable form also makes debugging and evolution practical. You can audit what the agent “knows”, adjust policies, or let the LLM reason directly over memory instead of treating it as a black box vector store.
Embeddings still make sense as an optimization layer, but making them optional rather than foundational is an important architectural choice if agents are meant to run long-term and stay coherent.
This feels less like a retrieval hack and more like actual infrastructure.
This is really helpful — thanks for sharing. If you don’t mind me asking, how long did it take you to build this? （I’m thinking about creating a small aggregation site to track interesting Agent projects and would love to get a sense of the effort involved.