A preprint submitted to arXiv on 12 November 2026 introduces Folio, a context manager that shrinks a long-running language agent’s active history while preserving access to captured evidence. The arXiv record dates the submission.
How recoverable context folding works
Folio links model-written turn notes, runtime-written event manifests and source records. The runtime folds existing views, showing older history at lower detail. Persistent handles support bounded reads, computation over stored text and a dossier of findings. The authors describe these operations in the paper’s system design.
What the authors measured
On 75 LOCA-bench tasks, the authors report 66 successes for Folio versus 54 for Piki-Classic, the same runtime using model-written compaction. All systems used DeepSeek-V4-Pro; this comparison used a million-token window. Each of five task chains ran once per system. These are preprint findings from the authors’ experiment.
LOCA-bench evaluates models together with their scaffolds as context grows. Its repository documentation describes local mock backends for services including Canvas, email and Google Sheets, plus records of agent trajectories and token usage. Those controlled environments define the setting behind the reported result.
The authors limit the evidence to one benchmark and model. Recoverability covers captured, retained records with valid handles; it does not guarantee correct evidence use. Model-written notes can also be wrong.
Why it matters for research offices
For researchers and universities assessing long-running agents, our reading is that Folio brings AI provenance into the memory discussion: an answer can be assessed alongside the record it relied on. A research office’s assessment would still need to distinguish retained evidence, accurate interpretation and a reproducible AI experiment. The reported task score addresses a narrower question than institutional readiness.
Sources
- Ruan and colleagues, Folio: arXiv submission record. Checked 2026-11-15.
- Folio v1: methods, evaluation and limitations. Checked 2026-11-15.
- Zeng, Huang and He, LOCA-bench preprint. Checked 2026-11-15.
- HKUST NLP, LOCA-bench repository documentation. Checked 2026-11-15.
