Anthropic’s undated research on AI memory offers a useful comparison for workers and small-business teams choosing assistants: in a small evaluation, saved notes helped most when the assistant could also search the original documents.
Here, memory means information saved for later sessions, not a change to the underlying model. A wiki is a collection of linked notes that the assistant writes and maintains.
The practical question is whether a reusable project summary improves answers enough to justify keeping it accurate. Saved notes can point an assistant to original records without replacing those records.
What the comparison actually measured
Anthropic tested 25 questions against roughly 350,000 internal files, developing the wiki over 128 sessions. It reports these shares of graded criteria met—not percentages of questions answered entirely correctly:
Information available | Reported score |
|---|---|
Document search alone | 53% |
Search plus developed wiki | 61% |
Developed wiki alone | 48% |
Calculated from those scores, adding the wiki to search gained eight percentage points; substituting the wiki for search lost five. This small vendor evaluation does not establish a general workplace improvement. Some other document tasks in the report showed no clear wiki advantage.
Start with the question you need answered
Search deserves a fair comparison. The ReFind research preprint reports competitive results by letting an assistant search original chat records without first generating a structured knowledge collection. Its comparisons reuse published results and do not hold every system’s search budget constant. That supports testing capable search—not declaring it the universal winner.
For a team choosing an approach, the following are practical recommendations, illustrated with hypothetical work tasks:
An occasional, exact lookup: To find the delivery date written in a particular purchase order, start with search and ask for the source. Add saved notes only if there is a recurring problem they can demonstrably solve.
Repeated questions across a project: To explain why a rollout plan changed across several meetings, consider a short, source-linked summary alongside search. Keep access to the meeting records so an exception or later change is not lost in the summary.
Preferences versus changing facts: A preferred report layout is a reasonable candidate for reuse. A project deadline needs a date and a source that can be rechecked. Do not treat both as permanently settled just because the assistant saved them.
The distinction between recalling a statement and working out what is current also matters. RohitAI’s explainer on tracking changes in long documents examines that separate problem.
Make corrections visible
As of November 6, 2026, Anthropic’s Managed Agents memory stores remain in beta. These are developer tools for building assistants, not evidence that every consumer Claude plan exposes the same workflow. Stored text can be inspected and edited through the developer interface or Claude Console. Changes create version records that let teams audit earlier content.
Anthropic’s Dreams feature is an access-request research preview on selected Claude API models. It reviews an existing memory store and past session transcripts, then writes a separate, reorganized store without changing the input. Despite the name, this is background note processing, not sleep or retraining the model. The documentation tells users to review the output before adopting it.
For a team, that separation is useful: compare a consequential correction with the dated source before allowing future answers to rely on it. A completed reorganization does not itself prove that every retained statement is true. Keep a clear way for people to report and fix a wrong note.
There is also a permissions choice. Memory stores default to read-write access; Anthropic recommends read-only access for reference material the assistant need not change. It warns that malicious instructions in untrusted material could otherwise persist in memory and affect later sessions.
Count the cost of keeping notes useful
Dreams charges standard API token rates for the selected model, and processing can take minutes to hours. Targeted corrections use the memory-editing interface; instructions to reorganize notes are not a substitute for an exact edit.
For a cost-only comparison, use the period before the notes become obsolete. Assume each future query saves a positive amount, with answer quality held equal. Total query savings must exceed the initial preparation and review cost plus the cost of updates. Include staff time as well as computing charges. This is an accounting framework, not a measured return from Anthropic’s experiment.
Quality needs its own check. The LongMemEval benchmark treats finding facts, reasoning across sessions, handling time, incorporating updates and declining to answer without evidence as distinct abilities. That is a useful basis for choosing trial questions, not a forecast of any current product’s reliability.
Before expanding a memory system, compare search alone with notes plus search on representative questions using the same model, source documents and scoring rules. Include a corrected deadline and a question the records cannot answer. Track review effort and recurring errors—not simply how much the assistant remembers.