Meta ProPerAgent Harness Lifts Proactive Agent Accuracy to 38%
The core insight is that timing, not memory volume, is the bottleneck for proactive assistants: a dedicated filtering call before the model sees a memory does most of the work. Builders of always-on or wearable agents can apply this retrieve-then-filter pattern without fine-tuning, and RL on top adds a further gain. It is a useful comparison point against standard RAG pipelines that flood the model with context.
Meta researchers present ProPerAgent, a harness for proactive agents that raises accuracy from at most 15% with standard memory baselines to 38% without changing model weights. The harness retrieves memories via a graph of past experiences, then uses a separate LLM call to filter whether retrieved memories justify intervening. Removing that filtering step drops accuracy from 37.3% to 24.0%, and unnecessary interventions on silence-requiring clips fall to 15.3% versus 77.3% or 84.0% for standard RAG. The harness reaches 37.5% with Qwen3.8-27B, and additional RL training raises accuracy to 42.4%.
Recommended work from Meta on harnesses for proactive agents.
The ProPerAgent harness raises accuracy from at most 15% with standard memory baselines to 38%, without changing the model's weights.
But this success requires both a useful response and correct timing.
The harness controls how personal memories reach the model as it observes the user's surroundings.
First, the model identifies objects in view. A retrieval module uses those objects to search a graph of past experiences, preferences, and constraints.
A separate LLM call then checks whether the retrieved memories justify offering help now. It filters out distracting or already resolved information and returns one concise piece of evidence, or nothing. The model uses that evidence and the current view to decide whether to speak or stay silent.
That filtering step matters. Removing it drops accuracy from 37.3% to 24.0% in an ablation. Across the full harness, unnecessary interventions on clips that require silence fall to 15.3%, compared with 77.3% or 84.0% for standard RAG in the two evaluation modes.
The harness also reaches 37.5% with Qwen3.8-27B without training. Additional RL training on that model raises accuracy to 42.4% by penalizing premature and repeated interventions.
For proactive assistants, the harness needs to check whether a memory warrants an interruption before passing it to the model.
Paper: https://t.co/fLvcHtmWpK
Chat with Paper: https://t.co/mexKsft27L
