Your AI assistant may not always have access to everything you have discussed. As a conversation grows longer, many tools compress the earlier exchanges into a summary to make room for new ones.
Most people find this out the same way: an hour in, it suggests something the two of you ruled out together, or asks again for a file you already gave it. Nothing has gone wrong with the tool. It may simply no longer see that part of the conversation, and it has no way of telling you so.
Here is what happens in between.
The one-page handover
An assistant can only hold so much conversation at once. As a session nears that ceiling, the tool writes a summary of everything so far and carries on from that. How much of the original survives varies by product — on Anthropic's API, everything before the summary is dropped outright. Your tool might announce this as compacting, or say nothing at all.
It works the way a shift handover works. The colleague who spent all morning on your problem goes home and leaves one page of notes. Whoever comes next is every bit as capable, reads the page, and picks the job up. But they only know what is on it, and they cannot know what was left off, because the thing that would tell them is the thing that is missing.
So your assistant is not ignoring the rule you set an hour ago. It never received it.
The obvious objection is that the tool should simply keep everything. That turns out not to help. Chroma tested 18 models and found the same pattern in all of them: the more you hand a model at once, the less reliably it holds onto any of it, even on easy tasks. A longer conversation is not a better memory. It is just a longer one.
Signs something was dropped
Nothing tells you what fell off the page, so the clues are in how the assistant behaves afterwards. Any of these can have another cause — a confusing prompt, a tool that failed quietly, a plain mistake. But when they appear in a long-running session, or right after you see it compact, a lost summary is the first thing to check.
Repeated work
One user's bug report on OpenAI's Codex, filed in March 2026 and still open, describes it well: after two or three summaries, the earlier decisions are gone, and so is any record of what has already been tried. An assistant circling back to an idea you jointly rejected is this in its clearest form.
Jobs taking longer to run
Researchers at JetBrains Research and the Technical University of Munich measured an assistant needing more steps for the same work when it ran from summaries instead of simple placeholders. Their explanation is the interesting part: a clean summary can iron out the very signals that would tell an assistant it is getting nowhere, so it carries on, agreeably, in the wrong direction.
Forgotten instructions
Anthropic's own warning is that summarising too hard loses details that are small but important. Whatever you said once, right at the beginning — use this format, leave that file alone — is the most fragile thing in the session, because nothing says it again on your behalf.
What you can do
You usually cannot turn the summarising off, and you would not want to — a conversation long enough to avoid it is long enough to hurt the answers. What you can decide is what makes it onto the page.
Keep records outside the chat
OpenAI's guidance for its own assistants is to keep the facts that matter in proper documents rather than a conversation due to be summarised. Decisions, rules, the current plan: put them in a file the assistant can reread, and the summary stops being the only copy.
Write the summary yourself
The person who filed that bug report found a trick worth borrowing: tell the assistant that every new summary must open with a short list of the decisions made so far. They report their sessions keeping the thread through four summaries and more. It is a user's own fix rather than an official one, but it costs nothing to try.
Check what it remembers
Before continuing a long session, ask the assistant to list the decisions, constraints and unresolved issues it still holds, then compare that against your own notes. Whatever is missing from its answer is what the summary let go. When you see a session get summarised, take one line to say the important rule again.
Start a fresh session
Once a session has been through two summaries, consider starting a clean one with a short brief and a pointer to your notes.
None of this stops the forgetting. It moves the decision about what gets remembered from the tool to you.
Working with Mandrill Tech
All of that depends on you being there to notice. An assistant serving a whole company has nobody watching. It summarises itself in the middle of someone's question, and the person waiting for the answer has no way of knowing anything slipped.
So a business assistant cannot keep what it knows in the conversation. It has to read from the company's own records every time it answers. That is the thinking behind Lestar, Mandrill Tech's AI finance assistant for Malaysian businesses: rather than pasting figures into a chat and hoping they survive, Lestar is connected to the accounting system you already use and answers from that data. How much cash is left, who owes you the most and for how long, whether the margin is shrinking — the answers come from your own books, not from whatever happened to still be in the conversation.
If you are planning an assistant for your business, talk to Mandrill Tech about what it needs to know, and where that information lives today.
