Memory should know which project it belongs to

One of the things I want most from an agent is not having to explain the same context every time we start working. I want it to remember the decisions that matter and pick up where we left off.

But which decisions should it remember for this task? That question matters more as I use agents across different kinds of work.

A preference for clear writing can travel with me. A procedure for checking a code change might help in several repositories. The reason one application uses a particular payment flow belongs with that application. If all of those end up in the same context, remembering more can make an agent less useful.

The difference shows up in the work

Suppose I have two projects. One lets customers pay when they book. The other sends an invoice after the work is complete. An agent helping with the second project should know how I like tests and reviews handled, but it shouldn't carry over the first project's payment rules.

It also needs the reason behind a decision. A step that looks redundant might exist because the business needs someone to review it. Without that context, an agent trying to simplify the code could remove the thing that makes the workflow correct.

That is why I think about several kinds of memory rather than one place to save everything. My preferences describe how I want to work. Project records hold decisions and current context. Skills capture procedures worth reusing. The research wiki gives me ideas and sources to draw on when a task needs them.

Remembering enough

I want a new session to find the relevant project history without reading every conversation I've ever had. I also want old decisions to be corrected when the work changes. A memory system that confidently supplies an outdated answer has missed the point, even if its recall is excellent.

This is one of the reasons agent memory keeps my attention. The interesting problem isn't just saving information. It is deciding what should be available, when it helps and when it should stop shaping the answer.

For my own setup, the test is practical: does the next agent start with a better understanding of this job, or do I spend the first part of the conversation untangling what it remembered?