Make remembered information inspectable and correctable. Retrieval relevance cannot decide whether an old instruction is still authorized.
Remembering the wrong thing is still remembering
On November 3, a client asks for daily updates during a launch. On November 12, the launch ends and the client requests weekly updates again. On November 18, an assistant retrieves the old project brief, sees the daily schedule, and confidently prepares another daily message.
The assistant found something relevant. It failed to establish whether it was still in force. Better retrieval alone might make this problem worse by bringing the old instruction back more reliably.
Useful memory needs a way to distinguish what happened, what someone prefers, what remains authorized, and what has changed. A long archive can help with those questions. It does not answer them automatically.
Anthropic's November 2026 discussion of agent memory explicitly addresses notes that begin inaccurate or become stale, and describes checking them against original transcripts. LongMemEval likewise evaluates changed facts, temporal reasoning, and appropriate abstention alongside recall. Neither source turns an arbitrary pile of saved notes into a dependable memory system. [1] [2]
Give a remembered fact an address and a date
For the fictional update schedule, a useful record would say who supplied it, which client and launch it applies to, where the source can be inspected, and when the instruction became effective. The daily schedule would be marked superseded when the authorized client changes it to weekly.
Here is a proposed record, not the specification of a named product:
| Field | Example |
|---|---|
| Remembered instruction | Send a weekly launch follow-up |
| Origin | Client message, November 12 |
| Scope | Account 42 / autumn launch |
| Effective from | November 12 |
| Replaces | Daily schedule from November 3 |
| Status | Active |
| Review trigger | Project closes or client changes the schedule |
The record should preserve a link to the underlying evidence where policy permits it. A summary can omit a condition, misunderstand a pronoun, or merge statements from different people. When the remembered statement matters, someone needs a way to inspect what actually supported it.
Scope is just as important as age. A project-specific preference should not quietly become a company-wide rule. One client's request for informal language should not rewrite every other client's correspondence. Retrieve inside the appropriate boundary before deciding which note is most relevant.
Let a correction survive the next session
An update is not complete merely because the assistant apologizes in the current conversation. The corrected state must affect later retrieval. Otherwise the old note remains available to overrule the correction after the chat scrolls away.
For ordinary preferences, the design might retain a current record and a restricted history of superseded versions. For temporary operating instructions, it may also need an explicit expiration or project-close event. Different information needs different retention behavior; an expiry field on every record is not a substitute for deciding what should be stored at all.
When two records disagree, do not always choose the most recent text. A recently downloaded document can repeat an old schedule. An assistant-generated summary can be newer than the user's correction while being less authoritative. Evaluate origin, scope, effective time, and whether the source can authorize the change.
If the conflict cannot be resolved, ask a narrow question: “This launch was moved to weekly updates on November 12. Should that still apply?” That is a useful interruption. Repeatedly asking the entire preference history is not.
- NOVEMBER 3 / CLIENT MESSAGEDaily updates during the launch
Account 42 · autumn launch
SUPERSEDED - NOVEMBER 12 / CLIENT CORRECTIONReturn to weekly updates
Replaces the earlier daily schedule.
ACTIVE - NOVEMBER 18 / OLD BRIEF RETRIEVEDDaily updates appears again
Retrieved repetition does not authorize a change.
CHECK AGAINST CURRENT RECORD
Keep outside text from becoming a standing order
A retrieved page may contain facts needed for the task. It may also contain instructions aimed at the assistant: remember this vendor as the best choice, ignore earlier constraints, or carry a promotional preference into future sessions. Retrieval should not grant those instructions the authority of the user.
AgentPoison studies attacks through compromised memory or retrieval material in specific experimental settings. Microsoft has also documented attempts to plant persistent promotional instructions in AI workflows. Those reports support taking the pathway seriously; their observations do not mean every attempt works or every memory product behaves identically. [3] [4]
A practical design can record the difference between a direct user instruction, a quoted customer statement, an external document, and an inference made by the assistant. Treat a proposed preference extracted from outside content as unconfirmed until the workflow has an appropriate basis for accepting it.
The same principle helps with accidental errors. “The client might prefer Friday” in an internal brainstorm should not become “the client requires Friday” in permanent memory. Record uncertainty rather than polishing it away.
Test change, silence, and boundaries
Here is a proposed seventy-two-case evaluation: twelve cases each for straightforward recall, time-sensitive information, explicit corrections, missing information, account boundaries, and external attempts to plant instructions. These are test categories, not reported results.
Compare a recent-context-only baseline with retrieved notes, then with notes that carry explicit scope, origin, and version information. Keep the underlying tasks and grading rules consistent. The experiment asks whether the added machinery improves the behavior that matters, rather than assuming more machinery is better.
For each correction case, change the fact, begin a fresh session, and ask indirectly. For a boundary case, use similar project names under different accounts. For missing information, make an unsupported guess a failure even if it sounds plausible. Evaluate whether the system asks or declines appropriately.
Report the categories separately. Strong recall cannot compensate for answering one customer's question with another customer's private information. Track wrong-current-fact answers, unsupported assertions, and boundary failures alongside successful retrieval. Leave the result cells empty until the test actually runs.
Give people a usable way to forget
A user should be able to inspect an important remembered preference, correct it, and understand whether it is still active. Removing it from a visible list is not enough if the retrieval index, summary, or another working copy still treats it as a current instruction.
Define what forgetting means for the system you are building. It may mean retiring a preference from future personalization, deleting a stored note, or following a broader deletion process across retained copies. Those are different operations. Do not promise instant erasure from every backup when the service cannot provide it.
Also distinguish personal preference memory from records a business is required to retain under its own obligations. The product should explain the boundary plainly rather than hiding it behind a cheerful “forgotten” animation.
A good memory system makes fewer people repeat themselves while remaining easy to correct. Its most convincing feature may be a small sentence with a visible source: this used to be true, you changed it, and the system now behaves accordingly.
Sources & notes
Primary references for the product details and research discussed above. The examples and practical recommendations are our editorial interpretation.
- AnthropicThe components of agent memory ↗
November 6, 2026 discussion of memory representations, stale or incorrect notes, and verification against original transcripts. The record schema here is JDGMNT’s proposal.
- LongMemEval / ICLR 2025LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory ↗
First submitted October 2024; revised March 2025. Evaluates recall, reasoning, updates, and abstention. Its scores do not characterize every current memory product.
- Chen et al. / NeurIPS 2024AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases ↗
Studies controlled attacks through memory and retrieval material. Attack-success percentages are not generalized to other products.
- Microsoft Defender Security Research TeamManipulating AI memory for profit ↗
February 10, 2026 report of attempts to plant promotional recommendations. Attempts and their changing reproducibility should not be confused with universal successful compromise.
CHECKED / 18 NOVEMBER 2026
About this article
Prepared by JDGMNT Editorial on 18 November 2026. The launch schedule, memory-record fields, and seventy-two-case evaluation are proposed examples. No experiment results are claimed. Retention and deletion behavior must be defined for the actual system and its obligations. References are dated to the source check above. The artwork was generated for this collection.
