Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-17 · high
The math ties how much a session keeps to squared prediction error, using Gaussian and nonparametric regression theory.
Why now: The preprint appears in the August 17 briefing on agent memory tooling.
A new paper splits AI context handover into three parts: verbatim decisions, compressed statistics, and raw observations, per arXiv 2608.14528. The split matters wherever a framework quietly truncates session history between context windows. Decisions stored as verbatim text can't drift into a different rule, and raw observations remain a fallback for whatever the statistics miss.
The model treats compaction and session restart as transferring an in-context learning state. It separates two goals summarizers usually blur together: recovering the exact original material versus preserving the distribution that material came from. Decisions and constraints get kept word for word. Repeated evidence gets compressed into statistics justified by the downstream task, not held raw.
Theory backs this with Gaussian linear regression and nonparametric regression, tying the memory budget to squared prediction error. That turns how much to keep into a number instead of a guess.
What I found most useful is the timing. A handover record gets written before anyone knows what the next session will ask of it. The paper prices that cost separately from answering the eventual query.
Gaussian and nonparametric regression are the only settings the theory actually covers, and the paper doesn't say how the bound holds outside them. Most agent memory tools I've used don't ship with any bound like that at all.
Each link below shares sources, entities, or timing with this story.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (compaction, constraint, context); pushes against this story (but).
Shared entity: Gaussian / Same source domain / Earlier coverage
Both cover Gaussian; reported by the same outlet (arxiv.org); earlier Gaussian coverage from 2026-08-04.
Shared entity: Theory / Same source domain / Earlier coverage
Both cover Theory; reported by the same outlet (arxiv.org); earlier Theory coverage from 2026-07-31.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (budget, context); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (budget, effect); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (budget, context); pushes against this story (vs).
Same source domain / Shared topic / Downstream implication
Reported by the same outlet (arxiv.org); overlapping topics (cost, decision); traces where this leads (downstream).
Reported by the same outlet (arxiv.org); overlapping topics (cost, downstream); traces where this leads (downstream).