Memory chaos in agents. Part 1: 2204 commands and zero entries

I asked the bug-hunter agent how many entries it had in memory. It said: zero.

A month and a half of work. 2204 terminal commands. Labels, donor sync, a spec with 55 fixes. And zero entries.

It wasn't being lazy. The folder for its memory didn't exist. For a month and a half it had been dropping notes into a box that wasn't there, and never once told me about it.

After that I went and checked the rest of them.

What I found

I measured by files, not by feel: 2.9 GB of chat logs, 167 memory sections, 289 entries. Seven projects, up to thirteen agents running at once.

Three agents had instructions that started with "read your memory". And named tools the server doesn't have. The ritual had never worked once in six months.

One of them was hunting bugs in code that doesn't exist. The instructions were written a year ago, half the folders are gone since then. It dutifully walked through nonexistent paths and reported back that everything had been checked.

The farm had 124 entries against a reasonable cap of 15. The table of contents showed 73. Fifty-odd entries were sitting invisible: taking up space, loading into every session, impossible to find.

Of 45 archived sections, 35 had never been opened once. Written and filed. Never read.

And now the worst part.

The most active agent had write access set to "everything". It could erase any section of any project — the farm, the store, the image generator. But its own instructions said it was restricted to its own section.

For six months it thought it was locked in its own room. And for six months it had the master key to the whole building sitting in its pocket.

Why this happens

Memory is tied not to the project but to the folder the agent is launched from. Give an agent a second working folder — it loses its memory. No error, no warning. It starts from a blank slate and doesn't even know something is missing.

This didn't pile up from laziness or stupidity. It piled up because the breakage stays silent.

A broken agent looks exactly like a working one. Just as cheerful in its replies, just as confident picking up a task. The difference is only visible if you open things up and look.

For six months I thought some of the agents were just slow.

Checking this yourself takes one question. Ask the agent: how many entries do you have in memory and where are you allowed to write. Don't read the instructions. Ask the agent itself.

In part two — how I sorted this out and what the numbers looked like.