journal / Published
Memory should know which project it belongs to
On August 25, I wrote about a distinction that matters in my agent setup: memory and rules solve different problems.
Retrospective. Event date: 2026-08-25.
Rules describe how an agent should work. Project memory holds context about the work itself. Sharing the rules can help. Sharing every project's memory with every agent can make things worse.
An assistant that answers questions across projects may need a wider view. An agent changing one codebase usually needs something narrower. It needs that project's decisions and constraints, not a large collection of unrelated history.
This is why I prefer memory with a clear scope. The goal is not to remember as much as possible. It is to bring back the right context for the task in front of the agent.
There is still a lot to test. More memory can be useful, but it can also preserve a bad assumption or carry context into the wrong place. I want to judge a memory system by the work it helps an agent do, not just by what it can recall.