Semantic Materialization: KV Cache Memory in Long-Horizon Agents
A recent paper on arXiv (2607.23693) investigates whether KV cache entries still hold value after the initial observations have been removed. The authors discover that by excluding one prior observation from the histories of agents, the cached rows of subsequent events provide a distinct view of computation despite missing inputs, a concept they term "semantic materialization." This phenomenon can be intentionally triggered: a well-crafted event that does not include an answer boosts donor-aligned recovery from 6% to 51% on Qwen3-8B without mentioning the target value. In contrast, passively collecting natural mentions from extended dialogues proves to be less effective. This study questions the validity of eviction and episodic-memory strategies in long-horizon agents.
Key facts
- arXiv paper 2607.23693 tests KV cache memory assumption
- Omitting one earlier observation from served agent histories
- Downstream cached rows act as independent view of computation
- Phenomenon called 'semantic materialization'
- Deliberate answer-free event raises recovery from 6% to 51% on Qwen3-8B
- Passive harvesting from long-term dialogue is less effective
- Challenges eviction and episodic-memory schemes
Entities
Institutions
- arXiv