Agentic Memory Fails Due to Prerequisite Eviction Before Retrieval

This research highlights a previously overlooked failure mode in agentic memory systems operating under fixed budgets: the eviction of crucial prerequisite information before it can even be retrieved. This phenomenon, termed "structurally indirect prerequisite eviction," occurs when upstream knowledge blocks, though weakly aligned with a current query, are discarded due to memory constraints, making subsequent retrieval impossible. The authors provide a clear operational definition of this failure and introduce a reproducible, deterministic benchmark to isolate and diagnose it. Through per-seed trace diagnostics, they demonstrate how this retention failure can severely impact an agent's ability to complete tasks. To address this, the paper proposes Dependency-aware Semantic Garbage Collection (DSGC), a rule-based approach that considers one-hop graph dependencies. Evaluations show DSGC dramatically improves full-chain retention, from 0.03 to 0.90 with a lexical encoder and from 0.23 to 1.00 with a sentence encoder in their main test suite. Robustness checks further delineate the budget and scaling regimes where this one-hop rule remains effective or begins to degrade, offering insights into the boundaries of this solution.