recursive-summarization-trades-detail-for-continuitymechanismmechanism reasoning
A running summary buys continuity across a long conversation at the cost of specifics, because it is lossy compression by construction — so it helps with "what were we doing" and not with "what exactly did I say in turn 40." Keeping the raw turns retrievable alongside the summary covers the second case.
Capability: Remembering across sessions · Context and memory, Chat assistant
Sources
- Reports the continuity half — recursive summarization produces more consistent responses in long conversations — and that the strategy complements retrieval-enhanced models rather than replacing them, which is the pairing this claim recommends.
- Separates long-term memory into distinct abilities including information extraction and multi-session reasoning, and finds gains come from indexing and retrieval design (session decomposition for value granularity, fact-augmented key expansion) — i.e. from keeping specifics addressable.
Disagreeing is the most useful thing you can do here. Both sides of every contested claim in this catalog were assembled by the same person, which is its weakest point.
Related claims
- When retrieval for a summarization task pulls definitions by surface term match from a general encyclopedia, ambiguous terms fetch the wrong sense and the model incorporates that unsupported background into its output, lowering claim-level factuality below the no-retrieval baseline even when the prompt tells it to ignore unrelated retrieved content.Stating false facts confidently · unreviewed
- Where a passage sits in a long input changes how much the model uses it — accuracy is highest when the needed information is at the very start or the very end and lowest when it is in the middle — so ordering retrieved passages to put the most relevant ones first is a real lever on accuracy.Losing information in long inputs · contested
- Writing out reasoning steps improves how a problem is decomposed but does not fix the arithmetic inside a step — the model can decompose correctly and still get the calculation wrong, at the same confident register as when it is right. For exact arithmetic the fix is to move the computation out of the text, not to add more steps.Digit-level arithmetic · contested
Notes
Backing is mechanism-reasoning, not single-paper, and the distinction matters here: neither source measures detail loss from summarization directly. The lossiness is definitional and the complementarity is reported; the quantified version of "how much is lost, and which kinds of detail" is not something I have a source for. Filed from the technique-efficacy backlog with that gap stated rather than papered over.