Answer
Why do AI agents forget what they were told?
Three different mechanisms get called forgetting, and each has a different fix. Telling them apart is most of the solution.
Three things get called forgetting: the session ended and nothing persisted, the instruction is still present but outweighed by later material, or a summarisation discarded it. The remedy differs for each.
The word forgetting imports an assumption that something was held and then lost, and only one of the three mechanisms works that way. Distinguishing them matters because the fixes are unrelated: a file, a restatement, or a change to how the session is managed. Applying the wrong one produces the common experience of trying several things and having the problem persist.
The first is the session boundary, and it is the simplest. Whatever was established in a conversation is gone when the conversation ends, because nothing wrote it anywhere. This is the case where the assistant genuinely does not know, and the fix is a file supplied at the start of each session. No amount of prompting inside a new session recovers something that was never stored.
The second is dilution, and it is the one most often misdiagnosed. The instruction is still present and is being outweighed by everything else that has accumulated, so it is followed approximately or not at all while remaining visible in the transcript. This looks exactly like forgetting and behaves differently: the fix is to reduce what is competing or to restate the instruction so it is recent, not to store it somewhere.
The third is summarisation. When a long session is compacted, or when a tool passes forward a condensed version, specifics are dropped by a process that does not report what it dropped. The instruction was held, then genuinely removed, and there is no trace of the removal. This is the mechanism behind the particularly disorienting case where the system contradicts something it said clearly an hour ago.
Diagnosing which one you have takes a moment. If it is a new session, it is the first. If the instruction is visible earlier in the same conversation and is not being followed, it is the second. If the conversation shows a compaction or the tool indicates the history was condensed, it is the third. Each points at a different action, and the diagnosis is what stops the cycle of repeating instructions more emphatically.
A fourth thing is sometimes mistaken for forgetting and is not: the instruction was never as clear as it seemed. An ambiguous rule is followed in one reading and then in another, which looks like inconsistency and is actually two reasonable interpretations of the same words. Before treating a recurrence as memory, it is worth reading the instruction as a stranger would.
Nothing was forgotten in the ordinary sense; either it was never kept, or it is still there and no longer winning.
Siddharth Sharma, Context Theory
Related questions
Does repeating an instruction more forcefully help?
It helps a little for the second mechanism and not at all for the first or third. Since dilution is the case where the instruction is present and outweighed, restating it makes it recent again, which does change its weight. In a new session or after a compaction, force is irrelevant because the words are simply not there.
Is this different from how a person forgets?
In a way that matters practically: a person who has forgotten usually knows they might have, and will say so or check. None of these three mechanisms produces that signal, so the output is equally confident whether the material is present, outweighed or gone. That absence of a signal is the reason the design has to compensate.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Three distinct mechanisms present as forgetting — nothing persisted past a session boundary, an instruction present but outweighed by accumulated material, and specifics removed by summarisation — and each requires an unrelated remedy. | Classifying an observed instance by whether the instruction is visible in the current session and whether a compaction occurred. |
| Software | Dilution is distinguishable from loss because the instruction remains visible in the transcript while being followed approximately, which means storage is the wrong fix and reducing competition or restating for recency is the right one. | Locating the instruction earlier in the same conversation and testing whether restating it restores compliance. |
| Response | Summarisation removes specifics without reporting the removal, which produces the case where a system contradicts something it stated clearly earlier in what appears to the user as one continuous conversation. | Checking whether the session underwent compaction between the original statement and the contradiction. |
| Constraint | None of the three mechanisms produces a signal of possible absence, so output confidence is identical whether the material is present, outweighed or gone, which is why the compensation has to be designed rather than requested. | Asking about material removed by compaction and observing whether the response is hedged. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one