Letting AI learn from experience is one of the hottest directions in AI agent development. A common approach is to have the AI record the "lessons" from each task, such as a rule, a summary or a technique, and apply them directly the next time a similar job comes up, so it improves without retraining the model. However, a research paper published on 19 October 2026, "More Memory, Worse Agents," points to a hidden problem with this approach: errors can be copied over and over through memory, making the AI worse over time. This article explains what an "error loop" is, the "Anti-Persistence" memory design the researchers propose, and practices businesses can borrow when managing AI memory.
How do AI agents "learn from experience"?
A large language model's parameters do not change while it is being used, so to let an AI agent accumulate experience, it is usually given an external memory. After each task, the system can store two kinds of content. The first is the raw record: which steps the AI actually took, and whether the task ultimately succeeded or failed. The second is abstracted knowledge: rules, summaries and techniques the AI has drawn from those records itself, such as "for this type of enquiry, check the order status first." When a new task comes in, the AI retrieves relevant items from memory for reference.
The benefits and risks of storing "lessons"
Storing abstracted knowledge has obvious benefits: the AI doesn't have to work things out again from a large pile of raw records each time, and can reuse a ready-made strategy, which is faster and cheaper. That is why many people assume "the more it remembers, the better." The problem is that the rules the AI draws are not necessarily correct. It may over-generalise from a handful of cases, or misread what actually caused a success. Once a wrong rule is stored, every time it is retrieved and reused, the error carries over into new decisions.
What is an "error loop"?
The researchers found that the damage goes beyond direct reuse. A wrong rule shapes what the AI actually does, and those actions become new raw records; when the AI draws lessons from those records again, it is likely to reach the same wrong conclusion. The paper calls this mechanism "error reproduction": errors are regenerated through a cycle of abstraction, execution and re-abstraction. In other words, even if you delete the original faulty rule, as long as the related action records remain, the error can be "learned" back again, so it persists far longer than simple reuse of one bad rule would suggest.
A simple example
Here is a hypothetical example for illustration. Suppose an AI customer-service agent, in its first few refund enquiries, happened to tell customers to wait seven days each time, and none of those customers followed up. The AI concludes a rule: "Always ask customers to wait seven days on refund enquiries." In reality, the customers may not have followed up simply because their refunds had already gone through automatically. From then on, the AI replies according to the rule every time, generating a large number of "please wait seven days" records; when it draws lessons again, those records reinforce the same rule. As a result, a practice that began as a coincidence becomes "experience" the AI firmly believes in, and it keeps affecting the customer experience.
The fix: Anti-Persistence
To address this, the researchers propose a memory design called Anti-Persistence, built on separating "accumulating experience" from "permanently storing lessons." Concretely, the system still keeps raw action records and outcomes, but no longer stores the AI's own guidance as permanent knowledge. Whenever a new task arrives, an adaptive selector looks at past outcomes to decide whether, and how, to put together guidance for that specific task on the spot; once the task is done, that guidance is discarded and not reused directly. The paper also acknowledges that because raw records are still kept, errors can still flow back indirectly through them, but direct reuse of a faulty rule is avoided.
Results, and what to keep in mind
According to the paper's abstract, across different benchmarks and different underlying models, Anti-Persistence achieved the highest overall success rate on held-out tasks and the lowest average cost in AI model calls among the memory designs compared. The researchers argue the results support separating experience accumulation from long-term storage of abstracted knowledge. Bear in mind, though, that this is a newly posted preprint that has not yet been peer-reviewed, and the results come from test environments set up by the researchers, so real-world performance still needs further validation.
Where do everyday AI tools store "lessons"?
This isn't only a problem in research experiments. The AI tools people use every day all have places that hold this kind of abstracted knowledge: the memory features in tools such as ChatGPT and Claude automatically record the AI's conclusions about your preferences; custom instructions and Project descriptions shape every later response; reusable Skills or technique files get applied again and again to similar work; and the knowledge base of an AI agent platform is the main reference an agent uses to make decisions. If wrong conclusions creep into any of these, they can affect the AI's performance for a long time, much like the error loop the paper describes. So it's worth regularly opening these tools' memory settings to see what the AI has actually remembered, and deleting or correcting anything questionable straight away.
Five practical ways for businesses to manage AI memory
Even if a business never builds its own AI agent, the research is useful for everyday AI use. First, keep raw records and outcomes: log what the AI actually did and how it turned out, not just the conclusions it drew. Second, label rules the AI summarised by itself separately from rules a person has confirmed; treat the former as "to be verified," and don't let them automatically become official policy. Third, regularly review the AI's memory, custom instructions and skill files, and delete anything outdated or questionable. Fourth, keep important business rules in a formal document maintained by people, as the AI's single source of truth, instead of letting the AI accumulate them on its own. Fifth, if you notice the AI repeating the same mistake, don't just fix the faulty rule; also check and clean up the related history, so the error isn't learned back again.
Conclusion: remember facts, not self-made lessons
More AI memory isn't automatically better; what matters is what gets remembered. Keeping facts about what happened is more reliable than keeping the "lessons" the AI inferred for itself, and the rules that really matter should be confirmed and maintained by people. Businesses that know how to manage AI memory can help AI staff keep improving instead of repeating the same mistake. To learn how to build reliable AI staff workflows for your company, visit ai.ud.hk to explore UD's AI Staff solutions.
懂AI，更懂你｜UD相伴，AI不冷