First frontier. Diagnosis found single-pass retrieval could not reach two-hop evidence, and the oracle mishandled partial cues.
Storage-time forgetting does not beat query-time selection
A pre-registered oracle-reader test of progressive memory demotion under a context-token budget. Because a perfect reader bounds any learned one, the hypothesis could be tested, and falsified, before training a single model.
Decision record
Written before the data. Checked against it after.
RULE · COMMIT 7459236
STOP unless progressive reaches the target with at least 1.2× less context than the best keep-everything baseline, at some redundancy level and target.
FALLBACK · COMMIT efc25df
Added after v2, before v3, for the case where no condition reaches the target. Not triggered.
OUTCOMESTOP
Every ratio below 0.56. The trained-reader phase was not run.
The result
Keeping everything and rendering only the query-relevant facts stays near ceiling down to 32 tokens. Progressive demotion plateaus near 0.67 and falls to the floor at small budgets.
−0.32 to −0.66utility gap at every budget · all 95% CIs below zero
8.9×fewer tokens per item with query-time rendering
Why it lost
Conflict questions, which ask where someone lived before, need the old value. Demotion had already stripped it. On those questions progressive memory scored 0.12, against 1.00 for query-time rendering.
Instrument history
Three runs. The label changed; the conclusion never did.
Oracle fixed via a monotonicity test; deterministic ties. Two-hop still failing, so no condition could reach the target.
Two-hop retrieval corrected. Every fix helped keep-everything more than demotion; the gap widened from about −0.15 to −0.33.
What this does not show
- Nothing about the KV cache, where the store is the context.
- One synthetic world family, heuristic policies, an oracle reader only.
- Query-time rendering was symbolic and charged at zero model cost.