Better AI memory starts with keeping decisions current
Useful memory needs clear current facts, visible corrections, and sources. We improved upkeep without introducing another service or database.
An assistant can retrieve exactly what you saved and still give you an outdated answer.
The search may be working perfectly. The problem is that the collection contains yesterday's plan beside today's correction, with no clear indication of which one applies.
After getting shared retrieval working, we focused on the quality of the records themselves. This stage required ordinary upkeep guidance and a practical correction test, rather than another memory service.
Give each topic a current answer
Long activity logs are useful for reconstructing history. They are harder to use as a quick account of what is true now.
Keep a concise current record for each important topic. It should explain the present decision, the reason it matters, what evidence supports it, and what remains unresolved.
Preserve history separately when it has value. Decide whether historical material belongs in the active search collection, and label it clearly if it does. Otherwise, a relevant old paragraph may compete with the corrected one.
This does not require deleting every past mistake. It requires making the current state easy to identify.
Make corrections explicit
Consider a fictional project record that originally says delivery is Tuesday. Appending “Friday might work better” leaves the reader with two possibilities and an uncertain decision.
A useful update says that Friday is current, Tuesday is superseded, who authorized the correction, and when it was made. Preserve unrelated facts while editing.
We tested that last detail by keeping a separate numerical value unchanged during a correction. Another agent then retrieved the new value, recognized the superseded one, and reported the preserved detail correctly.
That test did not establish that every future update would be captured automatically. It showed that the correction workflow worked in the case we checked. The full step-by-step method is in How to test AI memory without feeding it the answer.
Distinguish evidence from suggestions
“The owner approved this,” “an agent proposed this,” and “a public source claims this” have different meanings.
Save that distinction with the statement. A useful idea should remain a proposal until the appropriate person accepts it. An imported document should not be able to manufacture the user's approval.
Sources also need context. A price can be provisional. A test result can apply to one build. A successful demonstration can leave commercial readiness unresolved.
These qualifications make memory more useful because the assistant can reuse the fact without silently expanding its meaning.
Save outcomes after meaningful work
Good candidates include a confirmed decision, a reproducible fix, a failed approach with a clear reason, and a reusable lesson.
Temporary chatter, repeated status messages, raw tool output, and duplicate summaries often add less value to a current knowledge record. Keep detailed evidence where needed and link to it instead of reproducing it everywhere.
Before creating another document, check whether an existing topic record should be updated. If several contributors share it, preserve their unrelated work and check for intervening changes before saving.
Check freshness when it matters
A record that was accurate when written may no longer settle a question about current software, pricing, or availability. Use its date and source to decide whether fresh verification is needed.
During a task, reuse sufficient verified context where appropriate. Repeating the same lookup without a reason adds delay, while rechecking a changeable claim can prevent a wrong decision. The judgment depends on what changed and how much the answer matters.
Measure the benefit honestly
Clearer records may reduce repeated questions and conflicting answers. In our upkeep work, we verified correction and retrieval; we did not establish a speed improvement or run a broad reasoning benchmark.
Start with one confusing record. Make the current decision clear, attribute it, preserve what still matters, and ask the assistant a question whose answer you can check.
An understandable record helps the human owner review the system and gives the assistant better material to work with. That is a practical improvement you can make before adding more infrastructure.