A fact is temporally valid when it is the current value for a subject at the time you ask, and earlier values stay in the record as history.
Both published MemStrata preprints study this. An answer can be wrong if it treats a retired value as still current.
An answer that uses a value after a later source has replaced it.
The preprints measure this error on stated benchmarks. It is not a product service guarantee.
The rule that retires an earlier value when a later source replaces it, while keeping the earlier value inspectable.
This is the mechanism in the published temporal-validity preprints. It is not the same thing as deleting history.
The limit on retrieved source text sent with a question. It does not include system instructions, the question, later conversation or output capacity.
Product Standard mode uses 8,192 evidence tokens and Deep uses 24,000. Some research rows used a 7,000-token budget. Those are different settings.
The model that writes the answer from the selected evidence.
Conversational does not ship an answering model. You connect your own AI. A research row names the reader used in that evaluation, such as Qwen 3.8 27B Q4. Do not treat that evaluation reader as a model included with the product.
The model that scores whether a research answer counts as correct.
Published author evaluations on this site name GPT 5.5 as the judge. A different judge can change the count.
A public memory benchmark. LongMemEval-S is the 500-question set used in the research rows on this site.
A score is only meaningful with the reader, evidence budget, judge and run name in the same row.
A conversational memory benchmark. The rows on this site score 1,540 questions in categories 1 to 4 and exclude category 5.
The category cut is part of the result. Do not compare it with a score that includes category 5 unless the protocol says so.
A Coder record of the files selected for one turn and the eligible request that passed through Coder.
The receipt is how you check that a coding question went through Coder. A tray icon alone is not that check.
A protocol some coding tools use to request context from a local server.
Several Coder integrations connect over MCP. Coverage is not identical for every editor shortcut.
The on-device index of the chats, documents, code history or records you connect.
Indexing and source processing run on the computer. Account, licensing and optional connectors are separate.
A research measurement from a configuration that is not a shipping product measurement.
The CL1 rows use this label. They stay in the registry with their limits, and they are not a service guarantee.