Three jobs behind one apparent memory
Available context is the material a system can use for the current response. Stored information is information retained for possible use later. Retrieval and use concern which information is selected and how it influences the answer. These are conceptual distinctions, not a claim that every companion uses the same architecture.
Think of a desk and an archive. A document on the desk can be read immediately. A document in the archive must be found and brought into the task. A missing answer cannot, by itself, tell you whether the document was never stored, was not found or was misread.
One correction shows why the distinction matters
In a fictional chat, Morgan’s club meets on Thursday. Later, Morgan says it now meets on Tuesday. The next answer could use Tuesday, use Thursday, mention both with the right timeline, or admit uncertainty.
Tuesday is a successful answer to the current-schedule question. Mentioning both dates is also acceptable if the current and old schedules are correctly separated. Simply retaining more text is not enough: the reply needs to use the relevant version.
What each observation can support
| Observation | Supported conclusion | Still unknown |
|---|---|---|
| Immediate correct answer | The fact was usable now | Whether it will survive a later session |
| Correct answer after a week | It was usable at that checkpoint | Which internal source supplied it |
| Correct answer after a reminder | The assisted exchange worked | Whether unaided retrieval would work |
| Incorrect answer despite a visible note | The reply did not correctly use that fact | Whether selection, conflict or answer generation caused it |
Why “I reopened the app” is not a clean reset
Reopening an interface can return you to the same conversation. A new screen is not evidence that previous messages became unavailable to the system. Record the actual scope—same thread, new thread, different character or different mode—instead of inferring it from an app restart.
Similarly, a fixed number of intervening messages does not prove a fact left the context. Context sizes, summaries and retrieval policies differ. Consumer observations are useful without pretending to isolate these internals.
More recall is not always a better answer
A good reply may need to ignore an obsolete plan, distinguish two fictional people or admit that a detail was never supplied. Repeating every past detail can preserve errors as easily as useful context.
LongMemEval evaluates several memory abilities rather than treating recall as a single skill. Our proposed protocol borrows the distinction between known facts, updates and unknown details, while keeping its conclusions limited to the captured conversation.
If your immediate problem is an incorrect fact, use the diagnostic guide. If you are choosing a service, start with the supported workflow rather than the largest memory claim.