An agent can save your correction and still make the same mistake tomorrow. The note may be there. The missing step is bringing it into the next decision. We can put reminders in its context, add them at set moments, or let the model search for them. All three help. But a growing pile of notes does not, by itself, make the agent better at knowing which note matters now. That is the gap I want to look at here.

The note is there. Now what?

In the first article, I asked how yesterday’s experience changes today’s decision. Here is the part that bothers me: sometimes the agent only uses an old lesson after I point to it.

“We already discussed this. Check the note.”

Then it gets it.

But I did more than help it find a file. I made the connection between the current task and the old lesson. That was the part I wanted it to do.

A note on the shelf is not a note on the desk

Think of saved chats, notes, and Skills as books on a shelf. A Skill is a set of reusable instructions, sometimes with scripts or other files.

The model’s context is its desk: the request, instructions, and other information it receives for this step. It also has knowledge learned during training. Here, we are looking at newer experiences.

Retrieval brings something from the shelf to the desk. Using retrieved material to help generate an answer is commonly called RAG, short for retrieval-augmented generation.

But who picks the book?

If the reason to read a note is inside that unopened note, the system needs another clue to find it. An index, a reminder, or a useful description can provide that clue. Simply saving the note cannot.

Diagram of bringing saved memory into context: a short index first, then the useful notes
One way to bring saved memory into context: start with a short index, then open the useful notes

Experience changes what catches our attention

Think about looking for something in a familiar room. You may look toward its usual spot before you stop to recall the last time you saw it.

Your experience is helping direct your attention: what you notice and look at first. It is not just waiting for you to ask the right question. Research on human memory shows that learned locations can guide attention in this way.

In terms of the first article, episodic memory concerns past events; semantic memory holds knowledge. Both can shape what enters working memory, where we hold and use information now.

People still forget. But experience can change what we notice before we deliberately search. I want an agent’s experience to help it make that connection too.

Three ways to bring the past back

For an agent using a fixed model, saving a new note does not update the model’s parameters, the numerical settings learned during training. The harness, the software around the model, has to bring that note back into context.

Here are three common ways to do this. They can work together.

  • Put it there in advance. Files such as AGENTS.md or CLAUDE.md can supply instructions when the harness loads them. The model does not have to search for those instructions first. But receiving a reminder does not guarantee it will use it correctly.
  • Bring it in at a set moment. A hook runs when a chosen event happens, such as the start of a session or a tool call. It can fetch a reminder then. This gives us control over when the check runs. We still have to choose a useful moment and decide what the hook should do.
  • Let the model look for it. The model can call a search tool to find relevant history. This is flexible, but it has to decide what to look for—and sometimes whether to look at all.
Diagram of reminders loaded in advance, hooks at a set moment, and a search the model runs itself
These routes can work together. This figure shows hooks used to bring in memory; hooks can also run checks or enforce rules

First the index, then the details

Why not put every note on the desk? Old conversations contain unrelated, outdated, or conflicting advice. The context window, the amount of input a model can take at once, is limited too.

Progressive disclosure means showing a little first, then opening more when needed. For memory, that might mean loading a short index and reading the full notes only when they look useful.

A label like “Tuesday’s notes” tells us very little. A description of what the note helps with gives the model a reason to open it. If that clue stays buried in the full note, the model may pass it by.

We have made the shelf easier to browse. We still need to help the system choose the right page.

Finding it is not the same as thinking to look

I would separate two questions:

  1. Once the agent searches, does it find the right information?
  2. What makes it search for that information in the first place?

Better search helps with the first. Clear descriptions, reminders, and well-placed hooks can help with the second.

External systems can also use feedback to improve which memories they select. So the choice is not frozen forever just because the model is. But that improvement needs a process of its own. It does not follow automatically from saving more notes.

“When I point to the right lesson, I am still doing part of the remembering for the agent.”

More notes are not enough

I do not judge an agent’s memory by how much it saves. I care whether an earlier correction changes what it notices and does next. With an agent, we need to look at what gets loaded, what triggers a lookup, and what the model chooses to use.

External memory can help a lot. But a larger store of experience is not the same as better judgment about when to use it.

Bringing a lesson back is only part of the story. How did the agent turn that experience into a lesson—and what did it leave out? That is where the next article begins.

Up next in the series: How did the agent turn that experience into a lesson?

Resources

The article originally appeared on LinkedIn.