The Recall Debt: Flat Memory Schemas Structurally Fail Multi-Hop Retrieval

Long-lived AI systems store memory as a flat pool of documents and retrieve from it by lexical or semantic similarity. We identify a structural blind spot in this design — *concept bridges*: multi-hop chains whose terminal document shares neither vocabulary nor embedding proximity with the query. On an embedding-disjoint tail of the public MuSiQue benchmark ($n=159$ gold-annotated 2-hop chains), BM25 and dense retrieval recover the correct document **0% of the time** at recall@10. This is a failure of candidate selection, not reasoning: a long-context model answers 47.5% correctly when handed the terminal document, versus 5.0% from standard retrieval (+42.5pp). Nor is it a weak-retriever artifact — a stronger open embedder, a frontier API embedder, and late interaction all recover **≤22.6%**, far below the 61.6% reached once the single intermediate bridge is resolved. Measuring the same regime on a real operational corpus is its own trap: the standard mine-then-LLM-judge recipe yields a "concept-bridge" set that capability-grade judges find is mostly token-collision artifacts — a fresh $N=480$ audit puts the genuine-bridge rate at just **2%** in the disjoint tail — because embedding-disjointness conflates "bridged but distant" with "simply unrelated" and cannot serve as a validity gate. We call this accumulating, schema-level gap *recall debt*: the fix is a richer memory schema, not a stronger retriever. We release the anonymized, multi-judge-audited chain pack as supplementary material.