Bridge Starvation: A Reproducible Failure Mode of Single-Pass Memory Retrieval, with a Bounded Repair

Long-horizon agents can fail long before their final answer: a missed memory item at an early step quietly removes the evidence needed later. We isolate a reproducible instance of this problem in single-pass agentic-memory retrieval. The failure, which we call \emph{bridge starvation}, occurs when a query names entity $X$ but the answer requires a chain $X{\to}Y{\to}Z{\to}\dots$; the first retrieval pass finds the $X$ fact, but the bridge entities are absent from the query and the remaining links are never retrieved. We introduce \dkmp{}, a controlled trigger that varies chain depth (2--5 hops) and context length (1K--128K tokens), and we pair it with trace diagnostics that separate retrieval coverage from reader composition. As a bounded repair, \ours{} reuses the same retriever but asks the LLM to emit one short bridge entity between retrieval passes. On the targeted 2/3/4-hop, 128K-token regime, \ours{} reaches 1.00, 0.90, and 0.97 accuracy ($n{=}30$ per cell), within 3pp of the gold-support oracle on 3/4-hop and above IRCoT, ReAct, Self-Ask, and a passage-graph PPR variant under the same reader and index. The repair has clear boundaries: it does not solve 5-hop chains, where coverage falls to 0.75 and reader composition becomes a second failure mode; it does not help on short-context MuSiQue, where the trigger does not fire; and it does not target the cross-session aggregation failure that dominates \lm{}. We release the trigger, traces, and per-pass diagnostics.