Construction before search: a pre-registered factorial test of two repairs to LLM-extracted agent memory

A pre-registration for a factorial study of where the cost of agent memory is paid. In a production agent-memory system (nox-mem, a self-hosted service consulted by six agents), memory loses to naive retrieval over the same raw prose once the context budget is equalised in code: 42.4% against 63.6% answer accuracy, a net value of −21.2 pp, worsening monotonically with conversation size to −35.0 pp in the 200k–1M character band. Two defects of the write path account for it, and their effect is superadditive: repairing the construction window recovers 4.1 pp, repairing the extraction prompt recovers 4.1 pp, and repairing both recovers 21.2 pp. The plan registers that test before it is run. Its primary endpoint is the total effect of repairing both, which 500 instances can resolve; the interaction is a pre-registered secondary, and the plan states in advance that no sample this corpus allows powers it to the usual standard (78.1% at the full 500). It has two legs, separated by the evidence each needs. Production supplies that the window occurs naturally and its magnitude — observational, on the twelve sessions where the factor varies, reported as an existence result and no more. A public corpus (LongMemEval s_cleaned, 500 instances, all of them inside the character band where the damage was measured) supplies the interaction estimate with the window manipulated under our control, which gives both the sample size and reproducibility by third parties. What the document records against itself The plan is deposited with its own failures visible, because they are why it has the shape it has. Its first design (§4) was refuted by its own sampling code: the unit of inference is the session, production holds twelve with variance in the window factor, and eight were already spent. It is kept in the document rather than removed. An adversarial pass over the second design found two literal errors — the dose expressed in number of sessions where production measures characters of prose, and a truncation that kept exactly what the production mechanism discards — plus a power calculation that did not cover its own estimand and decision labels that contradicted their intervals. The dose was demoted from a point to a three-level factor after its denominator was stated: the median rests on six sessions whose seen-fractions span [0.020 .. 0.894], and 0.240 corresponds to no observation at all. The study was resized twice, and the first resize was half a fix: it corrected the variance and kept sizing the interaction test with the total effect as its effect size. Two adversarial voices from different model families found that independently on the same day. The decomposition is 4.1 pp + 4.1 pp of main effects plus a 13.0 pp interaction; detecting the interaction needs n = 522 against a corpus of 500, which is why it is no longer the primary. Independence was measured rather than asserted: 20.5% of session ids appear in more than one instance, but the coupling that can propagate into the estimate runs through evidence and touches 2.0% of instances, so the standard error is cluster-robust over answer-session groups. A second judge family brackets the criterion effect at [+8.1 · +21.2] pp depending on who judges, which is pre-registered as the reporting form. A positive interaction, on its own, does not identify the mechanism: two factors acting in series on whether evidence arrives and whether it is used produce superadditivity with nothing jointly repaired. That serial model is therefore the plan's substantive null rather than a caveat, and the production measurement exceeds it by 33× (+0.4 pp predicted against +13.0 pp measured), which is what makes the secondary refutable. A third family (GLM-5.3) then found that the plan's pre-set equivalence margin was unreachable at every sample this corpus allows — an interval inside ±5 pp needs n ≈ 2,100 against a corpus of 500 — so the plan no longer claims it can show absence, and states in the document why it cannot; the same pass removed two README claims the design never measured, for want of a selection arm and of an identifier-survival instrument. Every control the second design declared could be passed by a pipeline that returns the stored reference answer whenever an evidence session survives and refuses otherwise, without ever reading the constructed context. A counterfactual-evidence control was added: the fact inside the evidence session is altered, and the answer must follow the alteration. Not peer reviewed. The run it registers has not been performed. Files construction-interaction-prereg-v1.0.pdf: the plan. construction-interaction-prereg-v1.0.md: its source text. code-v1.0.zip: the measurement scripts the plan cites — the corpus census, the independence and evidence-coupling probes, the power calculation and the estimand check that supersedes it, the seen-fraction denominator — plus the PDF build and this deposit script. SHA256SUMS: checksums of the three files above.