Prepared 17 September 2026 as an independent Construct-2 ancillary study. Status: bounded pilot completed 17 September 2026. Preparation itself launched no runs; the subsequently authorized execution completed 108 task attempts.
Unchanged procedural guidance was the strongest short-horizon migration choice in this controlled SQLite pilot. After preserving a failed first acquisition screen, a compact bank maintained source correctness while reducing use tokens on fresh tasks. On twelve fresh tasks for Qwen3 8B, unchanged inheritance solved 12/12, no memory 1/12, and same-archive reconstruction 10/12. A policy chosen through paid recipient calibration also solved 12/12, but cost 83,495 tokens including setup versus 28,153 for unchanged inheritance. Matched source controls held facts and obligations fixed. This is a limited two-model, three-template result; longer-horizon payback and broader migration claims remain open.
Read the pilot report,
methods and prospective protocol,
development revision,
locked evaluation,
complete accounting, and
reproduction instructions.
All 108 submissions replay against their committed generators; raw requests,
responses, checks and failures are retained under runs/. The final
policy was locked at 092ae7d before evaluation. The local report closes this
bounded phase on explanatory progress, not publication acceptance or broader
question closure. Original PM1–PM3 expectations below are preserved.
Can paid experience with a previously unseen recipient improve the choice to retain, adapt or retire acquired procedural guidance enough to justify its cost, compared with competent unchanged inheritance and reconstruction?
The broader program asks how agents accumulate useful experience across sessions and where it should live. A lesson valuable to one model may be redundant, misleading or awkward for another. We want to understand that change in value, including cases where unchanged inheritance or retirement is the best choice. The investigator owns workload discovery, methods, execution and publication.
Establish functioning procedural guidance, then hold its source bank and underlying experience fixed while changing the receiving model. Initially keep facts and task obligations fixed. Compare inheritance with the recipient's own competent behavior; better absolute performance after an upgrade is insufficient evidence that inherited guidance helped. A harness change is a possible later extension, not a required second axis.
Useful alternatives include competent retrieval from the unchanged bank, the recipient without inherited guidance, reconstruction from common source experience, and a modest adaptation or retirement rule informed by paid recipient experience. Choose the contrasts needed to identify an explanation; a large factorial is unnecessary. Preserve ordinary tools, tests and feedback across alternatives. A target absent from model-identity labels needs a competent shared-bank fallback, not an artificially empty retrieval result.
Use developed guidance and serious alternatives before interpreting a negative result. Do not hide source experience from reconstruction or equate a rewrite with a useful adaptation. Count inspection, calibration, adaptation, validation, failed attempts and subsequent use. Distinguish source-bank construction from incremental migration, while accounting for both in a lifetime claim. Keep quality, tokens, latency and compute visible rather than concealing incomplete tasks in cost-per-success averages. Research search cost and deployment cost are different quantities; retain both where relevant.
Evaluate developed claims on fresh complete tasks after calibration. A post-hoc best action or accurate migration forecast is not an implemented decision rule. A simple policy, unchanged memory or retirement can resolve the comparison. Close a bounded phase on explanatory progress, a demonstrated limitation or a concrete resource constraint. No positive adaptation result is required.
- Memory Transfer Learning, 2604.14004v1, §§3.1–3.2/4.4, and CONTRAMEM, 2608.22533v1, Methodology and Appendix A.7, already support transfer comparisons. Include a strong unchanged bank; generic unseen-recipient transfer is not our contribution.
- MemCollab, 2603.23234v2, §§2.1–2.3, Appendix E, already uses model-aware selection. Our candidate concerns a recipient outside the bank's construction histories and the cost of learning what helps it.
- Does Your Agent's Memory Survive a Model Upgrade?, 2609.05339v1, §§2–3.2, Appendices C–E, already tests factual-store migration and cheap forecasting. This study concerns decisions about procedural guidance.
- Agent KB, 2507.06229v5, §§3.3/4.3, supplies feedback-conditioned refinement; Agent Memory Distillation, 2608.07169v1, §§3.2–3.3, supplies hierarchical construction. These are possible donors, not mandatory methods.
The root's versioned reading ledger
records inspected passages and limits. Agent-KB static inspection was at
588d6694e743a981179f2e11b47ca3a668e2f1ac; no author code was run, and that
snapshot is not verified as its paper's experimental revision. Consult the
upstream repository and pin any
artifacts actually reused. None of these outcomes has been replicated here.
Bounded feasibility work remains part of the commission.
Copied from the root's selection note before execution; preserve these statements when recording subsequent assessment.
- PM1: The useful value of inherited procedural guidance will depend on the recipient; some guidance that helps its source may be redundant or harmful after a model change. Uniformly useful inheritance would weaken this expectation within the tested scope.
- PM2: Paid recipient experience will improve adaptation or retirement decisions only where it distinguishes consequential alternatives. A strong unchanged bank, simple rule or observation cost may remove its advantage.
- PM3: Access to the same source experience will make competent reconstruction a stronger competitor than adaptation from the inherited artifact alone. Selective adaptation can still repay its cost when useful guidance survives and reconstruction repeats substantial work.
Start with AGENTS.md. In an assigned fresh session, develop the smallest consequential comparison, diagnose acquisition failures purposefully, and keep failures inspectable. Publish an abstract and links to methods, evidence and reproduction instructions here. Explain changes to the question as learning warrants them. This study can proceed independently of correction-lineage and executable-experience-retention; their unpublished work is not a dependency.