Agents that forget everything are exhausting. Agents that remember silently are worse. Operator’s answer is memory you can read.

What gets remembered

Preferences and facts about how you work, learned from what you actually chose. Not transcripts.

Things like: which model you pick for a kind of job, how you like reports structured, that you always want the tests run, a convention your team follows.

Where it lives

In Markdown files on your Mac, in plain readable English. Not a vector database you cannot inspect, not a cloud profile. Files.

There are two scopes:

• Per Solution. What the research jobs learned stays with research. A worker for that Solution wakes up knowing what the last one learned.
• Per person. The broader preferences, learned from your conversations.

Reading it

Every fact is listed, in words, in the app. You can read the whole thing in a couple of minutes, which is the point of keeping it small and in English.

Changing it

• Delete any single fact. It goes, and it does not come back unless you do the thing again.
• Delete all of it. A clean slate.
• Turn it off. Agents then start every job knowing nothing about you.

Turning it off demonstrably changes what agents do, which is a real test you can run yourself: turn it off, run a job you have run before, and see the difference.

Why it is worth leaving on

The compounding is the value. After a few weeks of using it, the jobs you run often stop needing the same three corrections every time, because the corrections became facts.

If you ever feel an agent is doing something you did not ask for, memory is the first place to look. The reason is usually written there in a sentence, and deleting that sentence fixes it.

What never becomes memory

Your API keys, your credentials, and the contents of your documents. Memory holds how you work, not what you have.