Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study investigates the implications of retaining key-value (KV) cache entries in long-horizon agents, specifically examining how the absence of certain observations affects the informativeness of retained events. The authors find that when an earlier observation is omitted, the responses are predominantly influenced by the omitted value, indicating a phenomenon they term "semantic materialization." Their findings reveal that the effectiveness of memory retention is highly dependent on the phrasing of events, with deliberate constructions significantly improving recovery rates in models like Qwen3-8B, suggesting a nuanced relationship between memory retention and the structure of information presented.
Omitting a single observation can dramatically skew agent responses, revealing that memory retention is more about how information is phrased than its inherent meaning.
Long-horizon agents increasingly reuse their KV cache as memory: a serving system keeps a subset of cached entries and drops the rest. Eviction and episodic-memory schemes therefore rest on a premise rarely tested directly, that a retained event is still informative once the observations that produced it are gone. We test it by omitting one earlier observation from what is served, across otherwise identical agent histories. Among items sensitive to that observation, the answer overwhelmingly follows the omitted value, though no served span says which value is correct. We call this semantic materialization: a downstream event's cached rows act as an independently servable view of computation whose inputs are gone. It can also be written on purpose. A deliberately phrased, answer-free event raises donor-aligned recovery from 6% to 51% on Qwen3-8B without ever naming the value, whereas passively harvesting natural mentions from long-term dialog yields no detected advantage. What such a row carries is specific and bounded. Compact state survives, larger payloads decay toward chance, and whether a construction writes at all turns on phrasing rather than on meaning alone, so two phrasings the model comprehends equally well can diverge sharply. The result is a memory contract for sparse event-KV serving: what to write, where it lands, and what survives once the source is gone. For anyone who evicts the corollary is that dropping a source event and observing no accuracy loss does not show the source was unnecessary.