Method & Playbook·7 min read

What an agent should remember (and forget)

The line between stateless agents and ones drowning in notes is ruthless curation of what actually persists across sessions.

By Andrew J. Pyle

The useful line between a stateless agent and one drowning in its own notes is ruthless curation of what persists across sessions.

Saving the transcript is not memory. It is bloat. Real memory distills: only operator preferences, feedback with its reasoning, project state, and resource pointers earn a durable slot. And forgetting is a feature, not a bug.

01DISTILL, DO NOT ARCHIVE

A transcript is not a memory

The easy, wrong version of agent memory is to save the conversation logs and reload them next time. That is not memory. It is bloat. It fills the context with everything and surfaces nothing.

Memory is distillation, not archival. The job is not to keep the transcript. It is to keep the insight the transcript produced.

02WHAT EARNS A SLOT

Four kinds of durable fact

Only four kinds of thing earn a persistent slot. Everything else is noise that should expire with the session.

KindWhy it persists
Operator preferencesHow I want work done. Stable, and expensive to re-learn every session.
Feedback with reasoningA correction plus the why, so the lesson transfers, not just the rule.
Project stateWhere ongoing work stands, when it is not derivable from the code or history.
Resource pointersWhere the real things live. The map, not a copy of the territory.

03CORRECTIONS COMPOUND

Feedback is the highest-value memory

Of the four, feedback is worth the most, because corrections compound across a fleet. One documented mistake, with its reasoning, stops every agent from re-learning the same error the hard way.

A rule without its reasoning is brittle. It gets misapplied the moment the situation shifts. Save the why, and the lesson travels to cases you did not foresee.

04WEED THE NOTES

Forgetting is a feature

Memory that only grows rots. A note that was true when written can quietly become false, and a confident wrong memory is worse than none. So forgetting is designed in, not neglected.

That means scheduled review, and verification before trust. If a memory names a file, a flag, or a number, it is re-checked against reality before it is relied on. A note is a pointer to be confirmed, not a fact to be assumed.

05NEXT STEPS

Memory turns a tool into a colleague

Without memory, an agent is a tool you brief from scratch every time. With the right memory, ruthlessly curated, it behaves like a colleague who remembers how you work and what you have already decided.

  1. Stop saving transcripts. Save the distilled insight instead.
  2. Admit only the four durable kinds: preferences, feedback, project state, pointers.
  3. For every correction, store the reasoning, not just the rule.
  4. Schedule forgetting, and re-verify any memory before you trust it.

Have something you need built or fixed?

I build production Django / Next.js platforms and human-supervised AI-agent systems. Solo, senior, and fast. Tell me what you are building.

Start a project