Field notes from building a memory layer for agents — the retrieval, the dedup decisions, and the architecture underneath.
Cosine similarity has no opinion about contradiction. When a user changes their mind, a pure vector store keeps both answers — here's why that breaks agents, and what a typed graph does instead.
Read postA look inside the async LLM pipeline that retrieves candidates, judges conflict, and writes one coherent memory.
Why we blend semantic and keyword signals with RRF instead of picking one — and how it survives messy real-world queries.
Every write is an event. How building Memuron on ArthaStore gives us audit, replay, and hard tenant isolation for free.
Occasional engineering notes when we publish. Unsubscribe anytime.