What does this research mean for the field?

Providing a canonical memory store outside the context window allows an agent to use handoff summaries purely for retention, enabling deterministic and lossless rotation of old material while maintaining high performance at reduced cost. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To evaluate whether separating window rotation from memory retention improves recall accuracy and cost efficiency in long-running autonomous agent sessions compared to standard summarization compaction.

August 30, 2026Open Access

pi-fold: Rotation and Retention

View Full Paper

Ask AI

Bookmark

Authors

SCShane Conner

Discussion

Member takes

Implication

Sequential case series shows external memory storage matches ceiling recall at half the cost in language model agents, suggesting rotation and retention should be decoupled.

Key Points

To evaluate whether separating window rotation from memory retention improves recall accuracy and cost efficiency in long-running autonomous agent sessions compared to standard summarization compaction.
Evaluated eight graded agent runs across a frozen 64-stage analysis of the curl codebase containing planted stale artifacts.
Tested a single model, provider, and adjudicator across compaction, lossless fold, self-annotated index, and external canonical memory store configurations.
Assessed agent recall using a withheld final exam querying co-occurring files within the conversation history that were absent from standard summary text.
Compaction scored 4 out of 16 on the final exam, with performance strictly bound by information transcribed in its handoff summary.
Lossless folding preserved raw conversation context but produced bimodal outcomes across four identical runs depending on autonomous recovery behavior, while agent index annotation yielded the lowest score.
Introducing an external canonical memory store enabled both compaction and fold mechanisms to achieve zero wrong answers, with the composed fold reaching ceiling performance at half the financial cost.

Cite This Study

Shane Conner (2026) studied this question.

synapsesocial.com/papers/6a93f0f56c1a8fb52e79d8a4 https://doi.org/10.5281/zenodo.22141606

View Full Paper

Ask AI

Bookmark

Implication

Sequential case series shows external memory storage matches ceiling recall at half the cost in language model agents, suggesting rotation and retention should be decoupled.

pi-fold: Rotation and Retention

Discussion

Member takes

Implication

Key Points

Cite This Study

pi-fold: Rotation and Retention

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider