Prepared 17 September 2026 as an independent Construct-2 ancillary study. Status: bounded feasibility experiments completed on 17 September 2026.
A fixed Qwen3.8 27B model acquired usable support links on one pilot and three prespecified fresh dispatch instances. All four methods—selective lineage, conservative exposure repair, correction-time reconstruction, and full rebuilding— repaired all persistent fields in 6/6 fresh episodes. All achieved complete state and later-task success in 4/6: shared read-time arithmetic errors caused the other two failures. Selective lineage recomputed 18 rather than 48 fields, but used 13,613 tokens versus rebuilding’s 9,789 (39.1% more), including ordinary writes, acquisition, maintenance, repair and audited use. It narrowly beat reconstruction over two corrections, while rebuilding remained cheaper. An unsupported maintenance link recurred on 1/3 fresh instances despite perfect initial link coverage. These are bounded synthetic observations, not a general lineage ranking.
Results and limitations · Frozen protocol · Evidence and cost ledger · Reproduction. The original commission and CL1–CL3 below are preserved. Preparation itself ran no experiments; execution was authorized in the subsequent assigned session.
Does acquiring and maintaining dependency information while memory is written improve later complete correction enough to repay its cost, compared with reconstructing dependencies or rebuilding from the same source archive?
The broader program asks how agents accumulate useful experience and where it should live. An agent can retain a conclusion and information about why it might need revision. The second investment may avoid later work, or cost more than reconstructing the dependency when needed. The investigator owns workload discovery, methods, execution and publication.
A runtime can record which inputs were visible or read. This is exposure provenance, not proof that every input supports the resulting memory. A model can propose selective semantic links, but those are acquired claims with possible omissions and spurious dependencies. Evaluator knowledge may score their quality; it must not silently become the treatment's supplied graph.
Begin with authoritatively supplied corrections or source invalidations while holding the receiving model fixed. This isolates propagation from fault detection. A later record is not automatically a valid supersession. Derived summaries, plans or procedural guidance may require repair, while independently supported conclusions and unrelated valid obligations should survive. Workload selection belongs here; no particular graph or benchmark is prescribed.
Develop competent alternatives: conservative all-input provenance, selectively acquired links, reconstructing dependencies from records when a correction arrives, and whole-state rebuilding or query-time auditing where feasible. Choose an identifying subset; combinations are allowed. Give alternatives the same retained source archive, current evidence, tests and checking opportunities. Withholding source evidence would confound lineage value with information retention. A gold dependency graph or randomly damaged gold graph alone cannot answer the acquired-lineage question.
Assess complete correction and later fresh tasks, persistent repaired state and unaffected obligations. A stale-item count is insufficient; erasing everything or refusing all tasks is not successful repair. Count costs during ordinary memory writes, dependency checks, withdrawal, reconstruction, repair, validation, later reads and failed attempts. Keep quality, tokens, latency and compute visible. Separate research search costs from a selected deployment's costs. Correction frequency and reconstruction cost are explanatory conditions, not invitations to tune until selective lineage wins.
- MemoRepair, 2605.07242v1, §§2–3.4, already studies dependency-based repair and edge deletion. A graph algorithm or generic missing-edge demonstration is not this study's contribution.
- Dependency-Guided Rollback, 2608.10502v1, §§3–4 and appendix schema, supplies selective recovery methods with runtime dependencies. Our comparison must account for acquiring useful semantic links.
- StateMem, 2608.19652v1, §§5.2–5.3, 6.2/Appendix H, already extracts dependencies and studies propagation. Mere acquisition is not novel; its lifetime value against reconstruction is at issue.
- StateAuditor, 2608.01619v1, §§3–4/6–7, motivates a competent source-grounded correction alternative. Do not compare retained lineage only with blind stale retrieval.
The root's reading ledger records exact scopes, reported metric limits and discovery boundaries. These methods are possible donors. No author implementation or experiment for these four methods was executed during preparation. Bounded feasibility and renewed overlap checking as the question sharpens remain part of the commission.
Copied from the root's selection note before execution; preserve these statements when recording subsequent assessment.
- CL1: Acquired selective lineage will save useful repair work only when its dependency judgments are reliable enough to preserve complete correction; precise-looking but incomplete links can make it worse than conservative exposure records.
- CL2: Reconstructing dependencies from a shared source archive will be competitive when corrections are rare or reconstruction is cheap. Maintained lineage can repay its write and checking costs when repeated corrections avoid substantial reconstruction.
- CL3: Preserving independently supported conclusions will matter for complete future behavior. Aggressive invalidation can improve a stale-memory diagnostic while damaging valid obligations, and a strong rebuild or query-time audit may outperform selective repair.
Start with AGENTS.md. Develop a small consequential comparison in an assigned fresh session. Diagnose unsuccessful acquisition of usable lineage purposefully; preserve failures and test developed claims on fresh material. Publish an abstract and links to methods, evidence and reproduction instructions here. A simple conservative method or reconstruction can be the useful outcome. Close on explanation, a demonstrated limitation or a concrete resource constraint. This study is independent of procedural-memory-migration and the active executable-experience-retention comparison; their results are not prerequisites.