Building a Local-First Memory Layer for AI Agents

The author has developed Signet, an open-source memory substrate for AI agents that aims to address the limitations of current query-based memory systems.

Details

The author argues that most current agent memory systems are essentially just query-time retrieval, which works when the user explicitly asks for something stored in memory but breaks when the relevant context is implicit. To address this, the author has developed Signet, an open-source memory substrate that handles memory outside the agent loop. Signet preserves raw transcripts, distills sessions into structured memory, links entities, constraints, and relations into a graph, and uses graph traversal and hybrid retrieval to build a candidate set, which is then reranked for prompt-time relevance and injected as context before the next prompt starts. This architectural shift aims to move from query-dependent retrieval toward something closer to ambient recall. Signet is local-first, inspectable, repairable, and works across various AI models. The author reports promising results on a small sample, with 87.5% answer accuracy and 100% Hit@10 retrieval.