Ask Only When Needed: Proactive Retrieval from Memory and Skills for Experience-Driven Lifelong Agents

The key architectural bet here is timing: most memory-augmented agent systems retrieve context before a task starts or reflect after it ends, but ProactAgent's ExpOnEvo mechanism triggers retrieval mid-execution, when the agent detects a knowledge gap it cannot resolve from its current context. That mid-task intervention is the specific claim worth scrutinizing.

This connects directly to the broader agent infrastructure conversation Modelwire has been tracking. OpenAI's updated Agents SDK (covered April 15) pushed native sandbox execution and long-running agent support, which implicitly assumes agents can manage their own context over time. ProactAgent is essentially proposing an answer to what that context management should look like internally. Google's 'Skills' feature in Chrome (April 14) is a consumer-facing version of a related instinct: making reusable knowledge accessible on demand. The research here is the more rigorous, lower-level version of that same problem, applied to autonomous agents rather than human prompt reuse.

Watch whether ProactAgent's benchmark gains hold when tested against long-horizon tasks that require more than three to four sequential retrievals, since the failure mode for proactive retrieval systems is typically compounding retrieval errors, not single-step accuracy.