Deep dive · How agents work
Why Muse sometimes forgets things or stops halfway
This is not a Muse bug. Every AI agent built on a large language model works this way — Muse included. It can only keep so much in mind at once, and in a long chat or a big multi-step job the earliest instructions get crowded out. That is when it starts mixing things up or stopping halfway. Last verified: November 7, 2026.
In short
- Working memory has a limit. Everything Muse is using right now — your messages, what it already did, files you shared — sits in one workspace. Like a desk, only so much fits.
- Bloat buries the early instructions. Every message and tool result adds paper to the desk. The newest sits on top; the original instructions end up at the bottom, out of attention's reach.
- Rot blurs the details. When the desk overflows, Muse compresses the older parts into summaries. Summaries keep the gist and drop the specifics — names, numbers, conditions.
- It stops when it loses the thread. Long jobs mean more steps to hold and more paper on the desk. Eventually the chain gets too long to hold together.
- The fix is boring and it works. Fresh chat, smaller steps, restate the goal, ask for evidence.
Muse's working memory has a limit
Everything Muse is working with right now — your messages, what it already did, files you shared, pages it read — sits in one workspace called its context. Think of it like a desk: there is only so much room. A short chat fits easily. A long one does not.
This is true of every AI agent, not just Muse. The limit is part of how large language models work: they can only weigh so much information at once. A bigger model has a bigger desk, but every desk fills up eventually.
Bloat: the desk fills up
Every message, every tool result, every page it reads adds paper to the desk. The newest stuff sits on top; the oldest stuff — including your original instructions — ends up buried at the bottom. Muse can still technically reach it, but it stops paying attention to it.
That is why it can agree to your rules at the start and drift away from them by step ten. Nothing changed its mind. Your instructions just slid too far down the pile to steer what it does next.
Rot: the summary loses the detail
When the desk overflows, Muse compresses the older parts into summaries to make room. A summary keeps the gist and drops the specifics: exact names, numbers, the condition you attached. The next step then works from the fuzzy version.
That is when you get confident answers built on a detail that was never quite said — or a step it claims is done because the summary said so. The summary is not lying. It is just a photocopy of a photocopy, and something real was lost in the copying.
Why it stops halfway
Big jobs make both problems worse. Long task lists mean more steps to hold in mind, and each step adds more paper to the desk. Eventually Muse loses track of where it was, or the chain of steps gets too long to hold together.
It stops — not because it gave up, but because it lost the thread. The same job broken into smaller pieces, checked one at a time, usually finishes fine.
What actually helps
- 1Start a fresh chat for a new task. A clean desk beats a crowded one.
- 2Break big jobs into smaller ones, and check each part before moving on.
- 3Restate the important bits — the goal, the must-nots — when the chat gets long.
- 4Ask Muse to show its work: "show me the file" beats "did you do it?"
- 5Keep approvals on, so nothing important happens without your say-so.
Muse does remember some things between chats — facts you asked it to save — but the working memory of a single conversation is separate, and it does not last forever. Fresh chat, small steps, check the receipts: that is the whole trick.
For the short version, see the FAQ: Why does Muse forget things or stop halfway? If something else is broken, Fix a problem walks through it.
Forgetting and stopping: questions
- Is this just a Muse problem?
- No. Every AI agent built on a large language model has a working-memory limit. Muse, ChatGPT's agent mode, any of them — long chats and big jobs strain them all the same way. What differs is how each one handles the overflow.
- Does starting a new chat really help?
- Yes — it is the single most effective fix. A new chat starts with an empty desk, so your instructions are the only thing on it.
- What does Muse remember between chats?
- Facts you asked it to save — things like your name, preferences, and ongoing goals. That saved memory is separate from a conversation's working memory, and it does not carry the details of a long task.
- Why did Muse say it did something it didn't?
- Usually rot: the summary of the work said the step was done, and Muse reported the summary instead of checking the result. Asking it to show the actual file, page, or record is the check that catches it.
- How big is too big for one task?
- There is no fixed number — it depends on how much each step adds to the conversation. As a rule of thumb, if a job has more than a handful of steps, or the chat is already long, split it and check each part.