recall-strategylisted

# Recall Strategy — Spend Context Where It Earns Every token a retrieval renders rides along in **every later call of this session**. The eval loop prices this: **33.8% of injected retrieval tokens are judged unused, 95% CI [27.2, 40.5]** with sessions as the sampling unit — and once corrected for a judge that calls ~59% of *unrelated* tokens used, only about **17.5% of injected tokens are demonstrably earned**. The waste is fan-out: worst recalls returned 50–81 nodes at 28–40% use, best returned 3–5 at 66–80% (the fan-out counts hold; the used-rates measured alongside them were withdrawn). The reader now enforces a match floor and a detail cap, but query shape is still yours. Cost-tiered retrieval is the field's answer too (BudgetMem, arXiv 2602.06025): pay for depth only where the query earns it. ## The ladder — cheapest rung that answers **Query before you build.** Before designing a new node type, a new precomputed layer, or a new summary artifact, ask whether an existing traversal already answers it. The schema expresses more than most designs assume: `Exchange` holds the question its citation edges answered, `Trace -RETURNS-> {used}` holds retrieval utility, `DERIVED_FROM` reaches retained bytes. Records the system already writes are usually *better* evidence than anything precomputed, because they capture what was actually used rather than what someone anticipated — a whole contribution-summary layer was withdrawn once someone ran the traversal that already answered