Anthropic Breaks Down Agent Memory: What Should Be Stored in an Enterprise Knowledge Base, and When Should It Be Referenced?

As companies expose AI to more and more data, a practical problem arises: Can we still trust the conclusions it drew last time? On November 6, 2026, Anthropic published a study on the composition of agent memory, discussing how to retain information gained during interactions for subsequent sessions and keep that information useful through organization and retrieval. Here, “memory” refers to external content that the model can access again; it does not involve modifying the model’s weights. (Original Anthropic Research Paper)

Memories are organized layer by layer from raw records

Research divides memory into several interrelated layers. Raw interaction logs preserve the sequence of events and can be reviewed when conclusions are in question; notes distill experiences and facts worth remembering at the moment; knowledge pages—or “Wikis”—organize scattered content by topic and link them together; and preloaded information brings a small amount of key content directly into a new session, allowing the model to access it without having to search first. (The Structure and Layers of Memory)

These layers are linked by organization and retrieval. The model allows users to review records across sessions, correct notes, merge duplicate pages, and then adjust which information should be loaded first. Retrieval is responsible for finding details when needed, while organization makes it easier to find them next time. A key design principle is that each layer should serve a distinct purpose, avoiding the repeated duplication of the same set of materials—which ultimately results in a change in format without reducing the burden of searching.

What evidence do the figures 61% and 53% provide?

The study developed a 25-question evaluation based on approximately 350,000 internal documents from Stainless. The search, combined with a gradually refined knowledge wiki, ultimately achieved a score of 61% on the evaluation items, compared to 53% when using search alone; the best result achieved by relying solely on the wiki was 48%. (Assessment Settings and Results) These figures support the approach of retaining search and adding organizational layers in this set of tasks; however, it cannot be directly inferred that any enterprise knowledge base would achieve the same benefits.

From a practical perspective, a wiki is useful for helping a system organize the relationships between concepts in advance, while search provides a way to access specific content. The volume of a company’s documentation, the frequency of updates, and the nature of the questions asked all influence the value of this combination. If answers to most questions can be found directly in a single document, you should evaluate whether adding a complex organization process is worthwhile based on your specific tasks.

Conditions for Preserving Memories

Information worth committing to long-term memory should include not only the conclusion but also the conditions under which it applies. For example, if a client requests a paper copy at a certain stage of a project, and the system saves this temporary arrangement as a permanent rule, it may be repeatedly applied in future work.

My recommendation is to document the source, scope of applicability, and last update date for any knowledge that could influence action, and to ensure that verified facts remain distinct from unverified hypotheses. A one-time ad hoc decision should not automatically be elevated to a long-term preference, nor should a single experience be generalized to all business operations in the absence of evidence. The quality of a knowledge system should be measured by its continued usefulness over time.

Conflict resolution is particularly important. When a new document supersedes an old conclusion, simply adding a note may still allow the old content to be retrieved. A better approach is to explicitly mark the replacement relationship so that the system knows which version is currently in use, while retaining the necessary audit trail. This requires the product to treat error correction as a routine function, rather than waiting until errors accumulate before cleaning them up.

Include maintenance costs in the revenue calculation

Organizing knowledge requires computational, storage, and maintenance resources. Businesses should evaluate the total cost over a given period: how much was spent on the initial organization, how much was saved through subsequent Q&A interactions, and what additional work is required when the information changes. Focusing solely on the number of tokens used for a single response can easily lead to overlooking the costs associated with the background organization process.

For practical testing, select a set of recurring business problems and compare three approaches: direct search, search with summaries, and search with structured knowledge. Record the quality of the answers, the accuracy of the sources, response times, and total cost, and conduct another round of testing after the data has been updated. If the new structure only improves the initial demonstration but fails to reflect changes in a timely manner, the more resources invested, the heavier the maintenance burden may become.

You can remember it, but you should also be able to delete it.

According to Anthropic's documentation on memory tools, file operations are performed by the application on the client side; the storage location and implementation are controlled by the developer, and measures must be taken to prevent operations from extending beyond the permitted directories. (Official Memory Tool Documentation) This serves as a reminder to enterprises that even after a model issues a write or read request, the application layer must still be responsible for boundary and data management.

If an employee loses access to a particular original document, can the summary previously extracted from that document still be viewed? When a user requests that a piece of information be deleted, are the associated pages, indexes, and automatically generated summaries updated accordingly? These issues must be clarified before the system goes live; simply hiding the original document from search results is not sufficient to consider the governance process complete.

Users should also have the opportunity to review and correct long-term records pertaining to themselves. Allowing the system to automatically organize data does not mean permitting it to unilaterally decide which information should be retained indefinitely. Different rules can be established based on business risk to determine which information expires by default and which conclusions require reconfirmation, ensuring that long-term data retention is exercised with the necessary restraint.

The insight this study offers for product design is to treat memory as a data system that is correctable, traceable, and reversible. Better-organized knowledge may improve responses, but it is sustained, effective source relationships and clear access boundaries that determine whether it can be effectively integrated into real-world business operations over the long term.