Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory
Kazuki Nakayashiki runs twelve registered studies and 14,760 attempts on one question: when an agent inherits terse memories and can pull only one archived source record, what form of directive written into the store actually steers that choice.
Ask this paper
A criterion beats a bare id by 35 points on six direct-provider models in Study D, length-matched, though the same contrast failed its registered superiority rule on a nine-model OpenRouter-served panel.
Appending the id cancels the criterion: on three Claude models the effect collapsed from 40/40 to 0/40 on Opus 5, and six byte-matched edits confirmed each exact string carries its own effect.
The cancellation is recoverable: a ratification line worth +96.0 points on Opus 5 plus a budget of two credits restored the target on all three models.
Model-dependent direction: continued into a decision, the criterion moved the choice toward the current record on Opus 5 (+100.0) and away from it on Fable 5.1, which is a warning against porting memory formats across models.
Unusually disciplined reporting: prospectively registered re-runs, explicit unresolved contrasts (fifteen of thirty within margin, fifteen unresolved, none beyond), and the author states these are descriptive effects of exact edits with no mechanism claim.
Abstract
An agent that inherits six one-line memories may pull at most one archived source record before acting; a directive written into the store can steer that choice: a pointer to the record, a criterion that identifies it, or both. Across twelve registered studies on one instrument lineage (14,760 attempts) we measured where the request goes under each form. On six direct-provider models a length-matched criterion exceeded a bare id by +35.0 points [+31.2, +38.8] (Study D); the contrast failed its registered superiority rule on a nine-model OpenRouter-served panel (Study E). Appending the id cancelled the criterion on three Claude models (Opus 5: 40/40 to 0/40; Study F-x); six byte-matched edits gave each exact string its own effect (Study G), and a re-run at eighty runs per cell left fifteen of thirty replication contrasts within the margin, fifteen unresolved and none beyond (Study G'). A ratification line (+96.0 points on Opus 5) and a budget of two credits restored the target on all three (Study J); across five criterion strings the suffix's cancellation held for four of the five wordings on Opus 5 and all five wordings on Fable 5.1 (Study H2); in a second store every model followed the criterion (Study H1). Continued into a decision, the criterion moved the choice toward the current record (+100.0 points, Opus 5) and away from it on Fable 5.1 (Study I). A one-character plan pointer's effect (+78.0 points; Study B, after a correction of its first repository report) returned the same verdict under a prospectively registered re-run (+81.7 points; Study B'). All results are descriptive effects of exact edits on fixed panels with registered intervals and no mechanism claim.