When sales teams ask an AI Agent to help prepare a quotation, they naturally expect it to remember the customer’s preferred currency and document format. However, if last month’s discount, an unconfirmed delivery date, or a temporary decision made during a meeting is also treated as a long-term rule, the next quotation may continue using information that is no longer valid. For businesses, an AI Agent forgetting something may simply require employees to spend extra time repeating instructions. But when it remembers something incorrectly, it can directly affect subsequent decisions.
AI Agent memory (agent memory) is the knowledge an AI Agent acquires while interacting with users and performing tasks. An AI Agent can retain this knowledge across work sessions and retrieve it later. Through a series of recent tests, Anthropic Research found that for memory to be useful, the key is to identify information that will remain relevant in the future and make it possible to review and revise that information later. If a business plans to use AI Agents for long-running tasks, it should first classify information into personal preferences, temporary states, and stable rules. It can then determine what may be written to memory, who should review it, and how errors should be corrected when discovered.
Why Does an AI Agent Need More Time to Organize Its Memory as It Stores More Information?
Anthropic Research’s game-based tests offer a concrete example. Claude Sonnet 3.7 took approximately 10,000 steps in Pokémon Red, accumulating 413 memory files and 225 folders, yet ultimately earned only one badge. Another model created just nine files and two folders in a similar test. These results cannot be used directly to evaluate enterprise performance, but they do highlight a practical issue: if an AI Agent records every change, it must later spend more effort organizing those records and determining which information is still valid.
In a business context, information can first be divided into three categories. Personal preferences are formats, tones, or working habits that a user applies repeatedly. For example, a sales representative may prefer quotation summaries to list the delivery date before the payment terms. This type of information may be worth retaining, but users must be able to view, modify, and delete it. Temporary states include current inventory levels, prices awaiting approval, and the status of customer complaints in progress. Because this information changes continuously, the AI Agent should retrieve it from operational systems or the latest documents each time it performs a task. Stable rules are tested, verified practices that will be used repeatedly, such as the sequence of quality-control assessments or the fields required for procurement price comparisons. Once validated by the responsible person, these practices can be organized into reusable skills.
A project team can begin by selecting one long-running task and classifying every item it plans to retain into these three categories. It should then ask: Will this information genuinely be needed the next time the task is performed? Could it change in the near future? If it is used incorrectly, who will detect the error and take responsibility for correcting it? If any of these questions lacks a clear answer, the information should not yet be added to long-term memory.
How Can Businesses Prevent an AI Agent’s Memory Errors from Being Reused?
Retaining information across work sessions is convenient, but it also introduces risk. When an AI Agent reads an old note during a later session, it may treat incorrect or outdated information as a reliable basis for action. In one test, Anthropic Research had an AI Agent compare its memory notes against the original transcripts. The AI Agent identified and rewrote an incorrect PokéCenter location. After the correction, it returned directly to Mt. Moon and eventually reached Cerulean City. In a test where the incorrect note was not corrected, the AI Agent retraced a previously traveled route three times. This test shows that screening information before it is written to memory is not enough. Businesses must also retain original records for future verification and conduct regular reviews.
During a review, the data source, scope of applicability, and last update time should all be verified. In a procurement comparison, for example, the statement “the supplier can deliver within two weeks” may apply only to a particular quotation, part number, and order quantity. If the AI Agent records only “two-week delivery,” it may incorrectly reuse that information during the next purchase. Unconfirmed assumptions should therefore be stored separately from verified information. When new data supersedes old information, the outdated content should also be modified or deleted immediately.
EgentWrX allows users to view, modify, and delete memories. Businesses can also configure different memory-writing methods based on the content: direct entry, review required, or approval required. Rules that must be applied consistently over the long term can be organized into skills. If a skill created by an individual employee is to be rolled out across the company, it must first undergo review. Implementation teams can begin by assigning owners for memories and skills, clearly specifying which information may be written directly and which information must be confirmed by sales personnel or process owners, and retaining the original records for future verification.
Should Company Documents Be Organized into a Wiki, or Is Direct Search Enough?
Memory, knowledge bases, and skills serve different purposes. Personal preferences can be stored in memory that users can view, modify, and delete. Company policies, product documentation, and forms are better retrieved through a knowledge base. Verified decision-making methods that must be applied consistently should be organized into skills. If these three are mixed together, one-time information may be mistaken for a company rule, and an AI Agent may fall back on personal memory when it cannot find the necessary data.
Anthropic Research tested a knowledge Wiki using approximately 350,000 internal files containing about 420 million tokens. The findings showed that a Wiki combined with a search tool was more useful for questions requiring information to be gathered across multiple documents, followed by judgment or calculation. If an answer could be found directly in a single document, however, a Wiki did not necessarily provide additional value. The organized content could even become larger than the original documentation. Businesses therefore do not need to organize every document into a Wiki before identifying use cases. Instead, they should begin with real business questions and determine the most appropriate way to retain the knowledge required to answer them.
For example, if customer service representatives only need to consult the latest return and exchange policy, direct retrieval should be tested first. If procurement personnel need to compare multiple supplier specifications, internal acceptance criteria, and historical records simultaneously, the business can then evaluate whether to use a Wiki. EgentWrX provides three types of knowledge bases: Wiki, vector, and tabular. It also allows businesses to set similarity thresholds and minimum citation requirements. If an AI Agent cannot find sufficient supporting evidence, it will refuse to answer. Project teams should first prepare questions encountered in actual work, specify the sources that answers must cite, and define the circumstances in which the AI Agent should refuse to answer. They can then compare direct search and Wiki results using the same set of questions and expand the implementation only after confirming that a genuine need exists.
FAQ
Should an AI Agent remember every user request?
No. Personal preferences that will be used repeatedly may be retained. Current prices, inventory levels, and processing status should be retrieved again when the task is performed. Verified rules that must be applied consistently should be confirmed by the responsible person and organized into skills, preventing one-time requests from becoming long-term practices.
How can a business determine whether information should be added to long-term memory?
First, confirm whether the information will still be needed the next time the task is performed, whether it is likely to change soon, and who will be responsible for correcting it if it is used incorrectly. If the source is unclear, the scope of applicability has not been confirmed, or the information is merely a temporary state associated with the current task, it should remain in the original records rather than being added directly to long-term memory.
What should a business do if an AI Agent has already remembered something incorrectly?
First, stop using the incorrect information. Then correct it by comparing it against the original transcripts, the latest documents, or system data, while also identifying any subsequent work that relied on it. Businesses should allow users to view, modify, and delete memories. Important information should also be reviewed or approved by designated personnel before it is written to memory.
Must company documents be organized into a Wiki first?
Not necessarily. If an answer is contained in a single policy document, direct search can be used first. If the task requires information from multiple documents to be combined and then analyzed or calculated, test a Wiki together with search. Before implementation, define the sources that answers must cite, the conditions under which the AI Agent should refuse to answer, and the person responsible for maintenance. This prevents knowledge content from accumulating without control.
References
- The components of agent memory — Anthropic Research’s study of AI Agent memory selection, review, and knowledge Wikis.



