How is this different from what a system remembers between sessions?

This is what is held while a task is in progress and it goes away when the task does. Anything retained across separate tasks is stored deliberately somewhere else, and confusing the two leads people to expect continuity that was never designed.

Why do long agent runs drift from their instructions?

Because displacement goes by position and the original brief was given first. It gets pushed out by twenty steps of accumulated history, nothing announces that it has gone, and the agent continues confidently on whatever remains, which is why the drift looks like a choice rather than a loss.

Summarising progress into a short standing account rather than carrying the full history, and restating the objective periodically so it returns to a recent position. Both are crude, both work, and both are the difference between an agent finishing and wandering.

Glossary

Working memory

What a system is holding while a task is underway, which is limited and is the first thing to run out on a long job.

In plain terms

Everything the system is currently keeping in mind: the instructions, what it has read, what it has already tried. It is finite, long tasks exhaust it, and when it runs out something has to go, chosen by where it sat rather than by how much it mattered.

Why it matters

Because the failure is quiet. Nothing announces that the earliest instruction has been displaced; the answers simply become less accurate in ways that look like the model getting worse, and the actual cause is a task that outgrew what was being held.

How it works

It holds the current task and nothing else. Instructions, retrieved material, previous steps and their results all occupy the same finite space, and every one of them competes with the others rather than having a reserved place.

Long tasks exhaust it, and agents make long tasks. Each step adds what was tried and what came back, so a process that runs for twenty steps is carrying nineteen steps of history by the end and reasoning about the last one through all of it.

What gets dropped is decided by position rather than by importance. The instruction that mattered most was frequently given first and is therefore the first to be displaced, which is why long-running agents drift away from their original brief rather than failing outright.

The degradation has no error. There is no message saying that something was discarded, so the symptom is answers becoming vaguer or contradicting an earlier decision, and the natural diagnosis is that the model is unreliable rather than that the task was too long.

Summarising as you go is the practical defence. Compressing what has happened into a short standing account, and carrying that rather than the full history, keeps the important part resident and is what separates agents that survive long tasks from ones that wander.

Retrieved material is the heaviest occupant and the easiest to trim. Pulling in ten documents to answer a question about one of them fills the space with nine irrelevant ones, so the quality of what is fetched matters as much for memory as it does for the answer.

Restating the goal periodically costs little and helps disproportionately. If displacement is by position, repeating the objective moves it back to a recent position, which is a crude fix and works for the same reason the problem exists.

What people expect to be dropped, and what is

This is worth understanding rather than memorising, because it explains a whole class of behaviour that otherwise looks like unreliability. An agent that performs beautifully for ten steps and then starts doing something adjacent to what was asked has not become less capable; the sentence describing what was asked is no longer in front of it. Once that is the diagnosis, the interventions are obvious and none of them require better models. Keep a short running account of what has been established and carry that instead of the transcript. Repeat the objective every few steps, which feels redundant and is not. Break work into shorter tasks with defined handovers rather than one long run, which also gives you somewhere to inspect. Each of those costs a little effort at design time and removes a failure that is otherwise extremely hard to diagnose, because the symptom appears long after the cause and looks like something else entirely.

Seen in the wild

An agent that drifts from its brief around the fifteenth step of a long job.
Lindy
A long conversation where an instruction given at the start stops being followed.
ChatGPT
A coding session where a constraint set early is quietly abandoned later.
Claude Code

Common misconceptions

People assume

Running out produces an error.

In fact

It produces worse answers with nothing to indicate why. Something is displaced silently, the output degrades gradually, and the natural conclusion is that the model has become unreliable rather than that the task outgrew what could be held.

People assume

The most important things stay.

In fact

What stays is decided by position, not importance. The instruction given first is frequently the most important and the first to be pushed out, which is exactly why long agent runs drift from their brief.

Questions

How is this different from what a system remembers between sessions?: This is what is held while a task is in progress and it goes away when the task does. Anything retained across separate tasks is stored deliberately somewhere else, and confusing the two leads people to expect continuity that was never designed.
Why do long agent runs drift from their instructions?: Because displacement goes by position and the original brief was given first. It gets pushed out by twenty steps of accumulated history, nothing announces that it has gone, and the agent continues confidently on whatever remains, which is why the drift looks like a choice rather than a loss.
What actually helps?: Summarising progress into a short standing account rather than carrying the full history, and restating the objective periodically so it returns to a recent position. Both are crude, both work, and both are the difference between an agent finishing and wandering.