Keep a Notes File Like an Agent

The best AI agents have one discipline in common, and it is embarrassingly low-tech: they keep a running plain-text notes file and read it back to themselves. They write down what they have done, what worked, and what to do next — then re-read it when their memory clears. The same habit makes humans reliable, and most of us abandoned it.

Anthropic's engineering team named the technique. "Structured note-taking, or agentic memory, is a technique where the agent regularly writes notes persisted to memory outside of the context window. These notes get pulled back into the context window at later times." ¹ An agent does this because it has to. It forgets. The lesson is that we forget too, and the fix is the same.

What most people believe about memory

Most knowledge workers believe their head is the workspace. You hold the project, the open threads, the half-formed argument, the thing you must not forget before the meeting — all of it in working memory, and you write something down only when it has already escaped. The notes file, in this view, is a backup for a mind that mostly works.

This is the model nearly everyone runs by default, and it is charitable to call it optimistic. It assumes attention is roughly free and that holding context in your head is the cheap, fast option. Steph Ango, the CEO of Obsidian, put the rebuttal in one sentence: "Apps are ephemeral, but your files have a chance to last." ² Your head is the most ephemeral app of all.

Why holding it in your head fails

The belief fails because working memory is small, and the machines we built to imitate us hit the same wall. Anthropic's engineers say it plainly: "Like humans, who have limited working memory capacity, LLMs have an 'attention budget' that they draw on when parsing large volumes of context." ³ A finite budget, spent on holding instead of thinking.

The agent designers reached for the obvious answer. "Like Claude Code creating a to-do list, or your custom agent maintaining a NOTES.md file, this simple pattern allows the agent to track progress across complex tasks, maintaining critical context and dependencies that would otherwise be lost across dozens of tool calls." ⁴ Lost across dozens of tool calls — the machine's version of walking into a room and forgetting why. The file is not a backup. It is the workspace.

What the research says about writing it down

The discipline holds up under the cognitive science, and the finding is sharper than "taking notes is good." Saving to an external store you trust improves your memory for the next thing, because off-loading the old frees the budget for the new. This is the human half of the agent pattern, measured a decade before the agents existed.

Benjamin Storm and Sean Stone tested it directly. "Results from three experiments showed that saving one file before studying a new file significantly improved memory for the contents of the new file." ⁵ The mechanism is the point: "saving provides a means to strategically off-load memory onto the environment in order to reduce the extent to which currently unneeded to-be-remembered information interferes with the learning and remembering of other information." ⁵ You remember more by holding less.

Storm, of UC Santa Cruz, stated the result in the open: "Our findings show that people are significantly better at learning and remembering new information when they save previous information." ⁶ And the broader frame: "By treating computers and other digital devices as extensions of memory, people may be protecting themselves from the costs of forgetting while taking advantage of the benefits." ⁶ A file on your device is not where thinking goes to die. It is an extension of the mind that does the thinking.

The effect is not an isolated lab curiosity. A 2022 review by Sam Gilbert, Annika Boldt, and colleagues defines the act precisely — "cognitive offloading, which has been defined as the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand (Risko & Gilbert, 2016)" ⁷ — and confirms the spillover: "saving one list of to-be-remembered words to an external store can improve memory for a subsequently presented list (Storm & Stone, 2014) or even performance of an unrelated task (Runge et al., 2019)." ⁸

One honest caveat, and it is also the most useful part. Storm and Stone found the benefit "was not observed when the saving process was deemed unreliable." ⁵ The boost depends on a file you can trust. A file in a service that could shut down or change its terms does not free the mind the same way; some part of you keeps holding on. So the case for an externalized file is also a case for owning it: a plain file on your own device, in a format that outlasts the tool.

The re-read is half the discipline

Writing it down is only the first move; the agents are reliable because they read the file back. A note is worthless if it is never reopened. Anthropic's worked example is a Claude model playing Pokémon over long horizons: it "develops maps of explored regions, remembers which key achievements it has unlocked, and maintains strategic notes of combat strategies." ⁹

Then the line that should change how you keep notes: "After context resets, the agent reads its own notes and continues multi-hour training sequences or dungeon explorations." ¹⁰ After context resets. Its working memory clears, and the first thing it does is open the file and catch up with itself. Most people's notes are written and never reopened. The agent's notes are a place it returns to. The difference between a graveyard of captured thoughts and a working memory is the second half of the loop.

What to do tomorrow

Keep one running plain-text file, and treat it the way an agent treats its NOTES.md — as the thing you write to constantly and read back before every reset. Not an inbox you fill and abandon. A workspace you return to. Five moves get you there:

1. Keep one file, not a system. A single plain-text or Markdown file you trust beats an elaborate vault you maintain. Start the file before you build the method.
2. Write the next action before you stop. End every work session by writing what you would tell yourself on the way back in. This is the agent's whole trick: a note from your past self to your future self.
3. Re-read it at the start, not just the end. Open the file first. The agent reads after every context reset; your reset is the next morning, the next meeting, the gap after lunch.
4. Off-load aggressively. Anything you are holding only so you won't forget it — put it in the file. The research says holding less is how you remember more. ⁵
5. Own the file. Keep it in a plain format, stored locally on your own device, so the saving process stays one you can trust ⁵ — and so the notes outlast whatever tool you wrote them in. Once the file exists, querying it is its own discipline. ¹¹

This is not a new productivity religion. It is the oldest one — write it down, read it back — handed to us a second time by the machines we built to think like we do. The plain file you would mention in passing turns out to be the architecture under the smartest software on earth.

Frequently asked questions

The questions below pair the agent-engineering pattern with the cognitive science underneath it, since the two halves explain each other. The short version: a stateless system, human or machine, stays reliable by writing to a trusted external file and reading it back — and the act of saving the old frees the mind for the new.

How do AI agents remember things between sessions?

Most language models are stateless: each session starts with a blank working memory. Agents get around this by writing to a persistent file outside the context window and reloading it later. Anthropic calls the pattern "structured note-taking, or agentic memory" — the agent maintains something like a NOTES.md and reads it back after a reset. ^{1 10}

What is agent memory?

Agent memory is how an AI agent carries context across tasks despite a finite context window. The simplest form is a plain file it writes to and pulls back in later: "Structured note-taking, or agentic memory, is a technique where the agent regularly writes notes persisted to memory outside of the context window." ¹ No database required.

Why keep a plain-text notes file?

Because saving to an external store improves memory for what comes next. Storm and Stone found that saving one file before studying a new one "significantly improved memory for the contents of the new file," by off-loading old information so it stops interfering with new learning. ⁵ A plain file is the most durable, portable form of that store.

Does writing things down improve memory?

Yes, and not only for the thing you wrote. Cognitive scientists call it offloading: "the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand." ⁷ Saving previous information measurably improves learning and recall of later information. ^{5 8} The catch: the benefit depends on the save being reliable. ⁵

How do AI agents keep notes in a NOTES.md file?

They write progress, decisions, and dependencies to the file as they work, then reload it. "Like Claude Code creating a to-do list, or your custom agent maintaining a NOTES.md file, this simple pattern allows the agent to track progress across complex tasks." ⁴ The same shape works for a person, reloading the file at the start of each session.

What is cognitive offloading?

Cognitive offloading is using an external action or tool to lower the mental effort a task demands — defined as "the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand." ⁷ Writing a thought into a file you trust is a textbook example: it frees working memory for the next problem.

The smartest software on earth was given one discipline — write it down in a plain file, then read it back — and it is the same discipline that makes a person reliable. Keep the file. Read it back. To keep your own running notes file in plain Markdown, on your own device, mnmnote.com is one open tab away.