The short answer
Do not rebuild the job from memory, yours or the model's. Rebuild it from the record. When work runs through the Zambo endpoint at https://zambo.dev/mcp, every call is written down outside the session as it happens: the step, its saved result, and its receipt. When the AI loses the thread, it reloads that record: the recent state of the job, step by step, exactly as run. It stops guessing what it did because it can read what it did. Zambo is the verifiable work layer for AI, and a reloadable record is what the layer is for.
The receipt, defined exactly: an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right. The record keeps the AI honest about state; judgment about the work stays with the reader.
The manual workarounds, honestly
Re-paste the brief. Scroll up, find your original instructions, paste them again, add "as I said before". It restores the goal. It does not restore the state: which steps ran, what they returned, which drafts were rejected and why. You end up re-litigating settled decisions because the AI cannot tell settled from new.
Keep your own running notes. A side document where you log progress by hand works, and it makes you the unpaid memory subsystem of your own tool. The notes also drift: the moment you forget to log a step, your record has the same disease as the model's context.
Start a fresh session and summarise. Sometimes right, and it throws away the one thing a long job accumulates: verified state. The new session inherits your summary, a retelling, and the forgetting starts again from a smaller base. Each rescue shrinks the job's reliable core.
All three workarounds share the assumption that the job's state is something a participant remembers. State should be something the job has, written down as it happens, readable by whoever is running it now.
The receipt-backed rebuild, step by step
- The job is recorded as it runs. Each call through the endpoint writes its step and saved result to the record outside the session, with a receipt. No extra effort during the work: recording is what the call does. You get 20 free calls per day to start.
- The AI loses the thread. You point it at the record. Instead of re-explaining, tell it to reload the recent state of the job: the steps in order, the result each one saved, the next steps as recorded.
- It re-derives the state, and can check it. The reloaded picture is not the model's reconstruction; it is the record, and any step in it can be verified at the public verify endpoint before the AI builds on it. Steps that already ran are known-good foundations, not rumours.
- Work resumes from the seam. The AI continues from the last recorded step, and its new calls join the same record, so the next loss of thread, in this session or a different AI entirely, costs one reload instead of one rescue.
The neighbouring shapes of this same problem: the session that ended outright is recovering work after a session expires, the model that forgets between one day and the next is forgetting between sessions, and handing the reloaded record to a different AI is switching AIs mid-task.
Why a bigger context window does not fix this
A longer window delays the moment; it does not change the architecture. The job's state is still stored as the oldest, most compressible part of a conversation, and compression is exactly where decisions and details get smoothed into plausible approximations. The record approach inverts the storage: state lives as structured steps on an endpoint built to hold it, in the AER-1 receipt format (draft-zambo-aer1-26), and the model's context holds only what it is actively working on. Forgetting becomes a read operation. The AI does not need to remember the job. It needs to be able to look the job up, and to check what it finds.
Questions, answered
Why does my AI lose track of a long task?
Because the task's state lives in the model's context, and context is finite and lossy. As the job grows, early instructions slide out or get compressed, summaries replace details, and the model starts improvising the parts it can no longer see. It is not carelessness; it is architecture. Any job long enough will outgrow the window it is being run inside, which is why the state of the work needs a home outside that window.
How do I remind an AI what it already did?
Give it something better than your recollection to read. If the job ran through the Zambo endpoint, tell the AI to reload its recent work state and read the recorded steps: what ran, in what order, what each step returned. That is a lookup against the record, not a retelling from you. If the job was never recorded, your only reminder is the transcript, and you are back to pasting and correcting, which works until the job outgrows that too.
Can an AI check its own earlier work?
Yes, when the earlier work carries receipts. Each recorded step's receipt commits to the exact saved bytes of its result, and the check is a recomputation at the public verify endpoint. So the AI, or you, can confirm that the step it is about to build on returned what the record says it returned. The receipt proves the saved result was not changed since it was saved; whether the result was right remains a judgment, now made against evidence instead of memory.
What is the difference between the chat history and the work record?
The chat history is a conversation: prose, drafts, dead ends, and the tool activity mixed into one scroll, owned by one app and one session. The work record is the job's state as structured steps: each call, its saved result, and its receipt, kept on an endpoint outside any session. A fresh AI can reload the record in seconds and know exactly where the job stands. Handing it the chat history is handing it a novel and asking for a status report.
What does it cost to keep a record my AI can reload?
The recording rides on the normal call terms: 20 free calls per day, then a $0.99 Day Pass for 24 hours of higher limits, or the Zambo Pass at $29 for 30 days, one payment, $290 for 365 days, and $19 for 30 days for $ZAMBO holders. Reading the record back and verifying receipts are part of the same tools, and verifying is free with no account. People pay by card through PayPal; agents pay in USDC on Base through x402.