pi-fold: Working Memory Under Context Shedding

Read full text
Authors: Shane Conner
Cited by: 2

Abstract

Long agent tasks can retain the assignment yet lose one transcript-only value required to continue. This report introduces a protocol-v4 stress instrument that plants seeded values only in the transcript, forces four decisions to be recorded during a 64-stage task, withholds a 30-cell block until delivery finishes, and classifies each expected value at its answering request. The evidence is an exploratory four-session sequential case series, not a fixed-N trial. Both pi-fold sessions completed. One native session reached stage 39 before an externally terminated harness runaway; the bounded native session stopped at stage 56. Across eight unique record-task endpoints, pi-fold recorded 8 of 8 correctly while native produced eight accepted unknown records. The two completed pi-fold sessions scored 40 of 42 ordinary probes and 60 of 60 withheld final cells after 5 and 9 context calls. At the answering requests, 53 final values were visible raw and 7 were recoverable but not visible; the seven actual answer carriers remain unidentified. For both native sessions, a read-only reconstruction exactly matched the sealed live post-compaction projection hash and found zero occurrences of the different continuation key needed next. These cases motivate preregistered replication. They do not estimate population rates, a native final-block contrast, a causal recovery effect, or a cross-arm cost ratio. The record includes a self-contained semantic HTML reading edition, five figures in SVG, PDF, and PNG forms, the relocatable schema-3 result, the frozen plan and seeds, the analysis-complete sealed evidence set, correction audits, exact source and Git history, the pinned workload archive, Pi 0.83 lock material, a privacy review, and SHA-256 manifests. Verification checks consistency inside the declared evidence boundary; it does not recreate the original provider interactions. Licensing note: the paper and author-created research artifacts are CC BY 4.0. pi-fold software in the source snapshot is MIT licensed. The included curl source archive retains its curl license.

Fields of Study & Subjects

Personal Information Management and User Behavior Mobile Crowdsensing and Crowdsourcing Human-Automation Interaction and Safety

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
17 August 2026
Document type
Preprint