Abstract
Persistent coordination memory can help multiple agents reuse prior experience, but a retrieved item may alter execution and therefore needs stronger authority than ordinary relevance ranking. We present FOG, a versioned executable-memory architecture that separates proposal, evidence admission, deterministic reduction, selective projection, and checkpoint rebase. We evaluate two stages under frozen synthetic protocols. For retrieval from an already constructed reminder topology, nine calibration cases are separated from 36 evaluation cases. Five retrieval paths are compared at result budgets 1, 4, and 8 under common lifecycle, scope, pointer, no-hit, and budget gates. All five pass the benchmark safety contract. A frozen lexicographic rule selects cue-graph retrieval, which at budget 8 attains macro target recall 1/1 with macro precision 25/196. A second protocol asks whether blank-slate execution receipts can form reliable constraint terrain. An untouched 16-seed held-out replication records fewer post-discovery repeat failures for induced FOG than for both controls at widths 4, 8, and 16. The frozen study verdict remains not_supported: the false-warning gate fails at widths 4 and 8, the precision gate also fails at width 8, and only width 16 passes every gate. Development and held-out results remain separate and are classified as other_divergence because their study verdicts agree but their per-width gate patterns differ. The evidence supports bounded topology-aware retrieval and records a width-dependent reliability pattern in terrain formation. It does not establish real-agent, language-model, production, or population-level effectiveness.