Memory is another document to maintain
An assistant helping a small team can remember the wrong thing perfectly. A decision discussed as a possibility becomes a settled plan in a note. The note is copied into a project overview, then appears at the beginning of every new session. Weeks later, the team spends an afternoon correcting a problem that began with one missing word: “proposed.”
Anthropic’s November 6 article, The components of agent memory, examines persistent notes, source transcripts, organized representations such as wikis, and context loaded into later sessions. It describes using offline processing to check notes against transcripts, reorganize information and repair links. The company reports benefits in several evaluations, alongside cases where a wiki offered little advantage or became larger than its source collection. These are results from particular tasks and setups, not evidence that every project should acquire a wiki or that background cleanup automatically produces accurate memory. The article treats this saved context separately from changes to model weights.
For a team using an assistant across sessions, the practical question is which remembered statements deserve to influence the next task. A short note can save repeated explanations. It can also conceal uncertainty or preserve a temporary arrangement after its reason has disappeared. The work is to review the revisions that turn past conversations into present assumptions.
Consider a fictional volunteer group preparing a monthly neighborhood puzzle newsletter. Its assistant helps keep track of page layouts, print deadlines and artwork. The group has no sensitive operations, and nothing in the example was tested. The project simply gives memory maintenance a visible purpose: next month’s editor should be able to find the current layout decisions without rereading every meeting. A good system should also let that editor discover when an apparently current decision is unsupported, conditional or out of date.
Compare the note with what actually happened
The newsletter’s memory note says, “Use a four-page layout for all editions.” Its source meeting transcript says, “Let’s try four pages for the September issue if the printer quote stays within our budget.” The note has dropped a date, a condition and the provisional nature of the decision. Reading only the note, an assistant could reject an October layout that the team had every reason to consider.
Ask for a small revision proposal rather than a general request to tidy the memory folder. The proposal should identify the current statement, the relevant source passage and the reason for changing it. In this case, the replacement might say that the team proposed a four-page September issue subject to the print quote, with the final decision not yet found. That wording does useful work: it preserves a lead without upgrading it into a standing requirement.
A transcript is evidence of a conversation, not necessarily evidence that the proposed action occurred. If the transcript contains an assistant saying the printer was contacted, that does not establish that the printer replied. If a participant says they think the quote arrived, that statement may still require the actual quote before the budget condition can be checked. Memory maintenance needs to preserve the kind of evidence as well as the content.
Use this distinction when the assistant finds a contradiction. One note may say the print deadline is Thursday while another says Friday. Before picking the newer sentence, inspect what each refers to. Thursday could be the deadline for artwork, Friday the deadline for submitting the final PDF. Both can be true. If the records really concern the same deadline, look for the person responsible for the decision and the scope of their message. Chronology helps find a change, but it does not by itself establish who had authority to make it.
When the evidence remains incomplete, retain the conflict as a question that affects the next action. The editor can ask the print coordinator to confirm the deadline. A short unresolved entry with a source reference is more useful than a smoothly rewritten sentence that hides the need to check.
Give the revision a place to land
For this small project, three kinds of record are enough to begin. Keep the original conversations or documents in a source collection. Maintain a concise page of current decisions with links to those sources. Put uncertain items in a separate list of open questions. This arrangement is a proposed method for the fictional newsletter, not a required architecture for every assistant.
Each current decision needs enough context to remain interpretable. The layout entry can name the edition, the decision, the person who confirmed it and the source where that confirmation appears. A temporary note can say when it should be revisited. Avoid adding fields merely because a template permits them. If nobody uses a field to find evidence or make a decision, maintaining it becomes another job.
Revision should change the current page while preserving an intelligible history. Suppose the coordinator confirms that the October issue will use six pages because a sponsor has supplied two puzzles. The updated record should make that the current October plan and mark the earlier four-page proposal as superseded for that edition. It should not erase the fact that four pages was considered, especially if a cost question later depends on why the plan changed.
Keep related changes together. Updating the layout page while leaving the overview at four pages creates a new contradiction during cleanup. A revision proposal can name every affected location: the edition page, the overview and the short note loaded at session start. The reviewer can then decide whether to change all three or remove a duplicate so there are fewer places to synchronize.
If several people or assistants maintain notes, give each proposed revision a clear owner until it is accepted. For a team this small, a visible pending-changes document may be sufficient. The objective is to prevent one writer from unknowingly replacing another’s correction. An orderly handoff is more valuable than an elaborate system that participants cannot explain or reliably follow.
Clean up the routes to the evidence
A memory collection can contain accurate facts and still be difficult to use. The newsletter’s overview might link to “printing plan,” while the relevant page is called “October production.” Another page may repeat half of the same information under “export checklist.” Searching for the deadline returns all three, and the assistant spends the beginning of each session deciding whether they describe separate tasks.
Treat navigation cleanup as a distinct revision. First check whether linked pages exist and whether the destination contains the information promised by the link. A link named “confirmed print deadline” should lead to the confirmation, rather than to a page that still asks someone to obtain it. This check is about the relationship between pages. It cannot establish that the deadline itself is correct.
Next inspect duplicates. Two pages may be redundant, but they may also describe different editions or different stages of production. Compare their scope before merging them. If the October checklist has a special requirement for a foldout puzzle, merging it into a general checklist could accidentally make that requirement permanent. A general page can link to the edition-specific exception instead of absorbing every detail.
Keep one place for the current answer when practical, and use references elsewhere. The overview can say where to find the deadline rather than copying its date. A person looking at the overview takes one extra step, but the team has one fewer stale date to repair. Whether that tradeoff is worthwhile depends on how often the answer changes and how often people need it. A frequently used, stable instruction may deserve a short summary.
Finally, check the effect on retrieval. Ask an assistant starting from the overview where the October PDF must go and what evidence confirms the deadline. This is a proposed check, not a claim that a tool has been tested here. A satisfactory answer should find the relevant decision and its support. Finding the right page by accident does not show that all links are sound, but a wrong turn gives a concrete cleanup task. Repair that route and repeat the question without expanding the entire memory collection.
Count the work of remembering
Memory costs more than storage. The assistant has to write notes, inspect them, follow references and revise them. Someone may need to review the proposed changes. For the newsletter, a growing wiki could become a second publication that demands attention while the actual newsletter waits. A smaller collection is not automatically better, but its size should have a reason.
Choose a few questions that memory is supposed to make easier. The editor needs to know the current page count, the artwork deadline, the required export format and who approves the final version. An assistant that can answer those questions with appropriate source support is doing useful work. A collection containing long biographies of every fictional volunteer may add little to that purpose, even if every biography is accurate.
Track maintenance separately from answering. Record the time or usage spent checking notes, rebuilding an index and reviewing changes, then record the work needed to answer the selected questions. If the team measures only the final response, it may overlook a lengthy background process that made that response possible. If nothing was measured, describe the result qualitatively rather than inventing a time saving.
Use a limit for the next cleanup pass. The newsletter team might choose to inspect only the pages needed for the upcoming edition, or only claims touched by recent decisions. A broader sweep can wait until someone finds evidence that the problem is broader. This keeps maintenance connected to an actual information need and gives the reviewer a clear stopping point.
Also consider what should disappear from the active view. A completed request to obtain a September illustration need not occupy the current dashboard forever. Move it to the edition record if the team needs the history. Remove a duplicate working note after its useful information has been incorporated and the team has checked what will be retained. Archiving something from active memory is a navigation choice; deleting the only supporting source is a different decision. Keep those actions distinct so a smaller dashboard does not quietly destroy the evidence needed to revisit a change.
Check whether the new memory earns trust
Before accepting a cleanup pass, compare the proposed result with a few questions that expose likely mistakes. Ask about the four-page proposal and whether it applied to every edition. Ask which deadline concerns artwork and which concerns the final PDF. Ask what remains unconfirmed. These questions inspect the meanings most likely to be lost during compression. They are more revealing than asking the assistant to summarize its own excellent organization.
Include a question whose answer should remain uncertain. For example, the fictional team has discussed a December color edition but has not approved a quote. The revised memory should help the assistant explain that status without treating the proposal as a commitment. If the cleanup turns every open item into a decisive answer, the collection may feel easier to use precisely because it has discarded necessary uncertainty.
Then check a new question that was not part of the revision brief. An editor might ask which decisions would need reconsideration if the sponsor withdraws its puzzles. A useful response can connect the October page count to the reason for it and flag the need for a new decision. It should not automatically revert the issue to four pages, since that action was never authorized. Good organization helps recover dependencies; it does not give old notes authority to choose the next plan.
Keep the previous version available while the team reviews the change. If a merged page loses a necessary exception, restore that passage or revise the merge. Record the cause in a concise maintenance note so the next cleanup can avoid the same mistake. There is little value in accumulating a large narrative about every editing choice, but a repeated failure deserves a visible explanation.
The final test is practical: the next editor can locate the current decision, follow its supporting evidence and see what still needs a person’s judgment. A memory system that does this may remain quite small. It may use a handful of notes, a source folder and an index. What gives it value is the checked path between them, plus the willingness to revise that path when a confident note turns out to remember more than the source ever said.
Try this, then make it yours.
Pick one non-sensitive project note and compare it with the conversation that produced it. Look for a missing date, condition or decision owner. Propose a correction with a source reference, then check whether the overview and related notes would also need to change.
Follow the signal.
Our reporting starts here. Practical suggestions are our analysis, and vendor performance statements are claims unless independently verified. We haven’t hands-on tested this release.
- Anthropic: The components of agent memory ↗ · 2026-11-06
