I've been experimenting with a simple idea: what if coding agents had durable, repository-local memory instead of relying only on the current chat context?

I open-sourced an early implementation called Agent Memory Layer.

The goal is to preserve project intent, important decisions, and structured knowledge inside the repository so AI agents can maintain continuity across sessions. The vision is for the AI to be the primary producer and consumer of this memory, with humans reviewing when needed.

Repository:
https://github.com/ragnarok268/agent-memory-layer

I'm looking for real-world feedback from people using AI-assisted development. Does this solve a problem you've encountered, or do you think existing approaches are sufficient?