AI Agent Memory: You Probably Have A Filing Problem.

Ryan “RyMac” McKinney · USMC · 15 August 2026 · 6 minute read

AI agent memory is the ability to carry information across sessions so the agent does not start from zero every time. It is a real capability and there are real tools for it. It is also the thing people buy when the actual problem is that nothing was ever written down anywhere the agent could look.

Two different failures that feel identical

Failure 1: it forgot. You told it something. It knew. Later in the session it did not. That is not memory, that is the context window filling up and pushing things off the desk.

Failure 2: it never knew. The information lives in your head, or in an email, or in a document you did not point it at. There is nothing to remember.

Memory tools address a third thing: carrying what it learned into tomorrow. Useful. Neither of the 2 failures above.

The test

Open a brand new session. No history.

Ask it where one of your projects stands.

If it finds the answer in 2 moves, your filing is sound and a memory layer will genuinely add something on top.

If it cannot, memory will not save you. You would be buying a system to remember information that was never recorded.

Most setups fail this. Mine failed it the first time.

Why it rots at week 3

This is the pattern I keep seeing, and an agency owner said it better than I can:

Nothing broke. The setup was carrying state in the conversation rather than in files, and conversations end.

Every session started slightly further from the truth than the last, and by week 3 the drift was visible.

A folder does not drift. Not a clever argument, just what happens when the state is plain text you can open and read. When something is wrong you can see it, because it is a sentence.

What to do before you buy a memory tool

Write down the 4 things that change. What we are working on. What is already decided. Where the material lives. What happens next.

Put them in 1 file. Have the agent read that file first and update it last.

~/business/
├── CLAUDE.md          <- the map. 11 lines. always read first
├── Context.md         <- the desk. what we're doing today
└── ops/
    ├── clients/
    ├── offers/
    ├── sops/
    └── archive/

That is persistent memory. It is a text file. You can open it on a laptop from 2011 and read it yourself, which is a property no memory service offers.

When a memory tool is genuinely right

I am not against them and I want to be fair.

If you need semantic recall across thousands of past conversations, or you are building a product where users expect the assistant to remember them personally, that is a real engineering problem and a folder is not the answer.

The question is whether you are that, or whether you are somebody who has never written down what the business is doing this week.

Where the method comes from

Interpretable Context Methodology, ICM, the paper by Jake Van Clief and David McDermott. Breakdown here: Interpretable Context Methodology in plain English. Wider concept: context engineering. Jake teaches it himself over at Clief Notes.