Solving Agent Memory Problem

I’ve been experimenting with solving lost context and evolved my system into a custom plugin harness to ensure queries read/write to the memory store. On the positive, I run into far fewer issues of lost context and can swap terminals and pickup work. But it’s not consistent and I don’t feel like it’s an optimal solution yet. I’m exploring other approaches.

Here’s my approach:

Memory: I have a central memory store with short term and long term memory. Short term stores context for active work. Long-term captures aggregate data accross repos with SQLite. My plugin reads longterm at start of a project and updates at end. Short term memory is checked before each task or request.

Plugin: routes Claude Code/Codex requests through a 4 stage loop: assess/plan/execute/review. Assess state pulls context and review stage writes back new info. Additional retrospective, context, and lessons learned captured after each push.