遇见数据集

Does the form in which a status reaches agent-memory writers change what a memory keeps? Protocol and materials frozen before the run (Study 5b)

收藏
Zenodo2026-10-30 更新2026-11-05 收录
官方服务:

资源简介:

Frozen protocol, code and materials of a follow-up study (Study 5b) of whether the form in which a final status record reaches the memory-writing configurations of LLM agents changes what their memories keep, deposited before any smoke or confirmatory model call (the manifest's OpenTimestamps proof is included; its Bitcoin attestation is checked by the gated launcher). On scripted histories of a recurring operational task in which the standard route is unavailable for maintenance and a final status record says whether it is still unavailable or available again, the final status reaches the writers either as a session of its own (a status message and the agent's acknowledgement, as in the earlier Study 5) or inside an operational session, a scheduled status check whose tool call returns the status; the scheduler record a reader is shown is the same in both forms. Four memory-writing configurations (a replay of the v1 memory pipeline of OpenAI's open-source Codex agent with its default writer models; ReasoningBank's released extraction instructions; mem0's open-source pipeline 2.0.19; Graphiti 0.30.2 with an embedded graph backend) write memories from 256 histories (64 families, both versions, both forms); fresh sessions of eight models choose the route for the next run after the record, after a memory and after a memory beside the record, and each choice is scored against the route that the history's stated rules and costs require. The registered primary contrasts compare, per configuration, the ended-case errors after the memories written from the two forms. The archive holds the registration (paper67/study5b/REGISTRATION5b.md), the design document, the decision log at the freeze, the study's code (which executes the frozen files of the earlier Study 5 and of the author's Paper 7, unchanged in their bytes and checked against their freeze manifests), the pre-probe commitments with their timestamps and Bitcoin attestations, the development probe's gates, reports and run manifests, the registered calibration of the statistics and its report, the registered histories, the pre-freeze reviews with their dispositions, and the chain scripts. MANIFEST.sha256 lists every file with its SHA-256; MANIFEST.sha256.ots is its OpenTimestamps proof. Manifest SHA-256: eb7fdc78230e8b972ae92e2f6e69f2df1e9191d724b693a25a96a81ecbec3aea. Zip SHA-256: 5d937a1b49f8a573a5b2a212467f5f2b55a1744f16abe7156c74a4e8dd946a1f. No confirmatory result is included. Licences: paper67/study5b/LICENSE.md.

提供机构:
Zenodo
创建时间:
2026-10-30
二维码
社区交流群
二维码
科研交流群
商业服务