Meta's ProPerAgent lifts memory-driven proactive assistants from 15% to 38%
dair_ai · x · 2026-10-20
Meta (with MSU) released ProPerAgent: Toward Memory-Driven Proactive Assistance, studying assistants that observe a user's surroundings (egocentric video) and decide on their own when to speak and what to say.
Key findings:
- Their new benchmark ProPerMem pairs egocentric video with year-long user memories (618K tokens avg); all baselines score at most 15%, including frontier models
- With oracle evidence GPT-5.3-Chat hits 96%; replacing it with retrieval drops to 38% — content and timing both matter
- Removing the filtering step drops accuracy from 37.3% to 24.0%
ProPerAgent harness: the model identifies objects in view, a retrieval module searches a graph-organized memory store of past experiences/preferences/constraints, and a separate LLM call gates whether the memory justifies intervening now. Without any training it lifts Qwen3-27B and GPT-5.3-Chat from under 10% to 37.5% and 38%; a lightweight RL stage reaches 42.4%. Code and eval set to be released.
More from coding & agent
- Dev turns his Goodreads shelf into a slick site you can rebuild with Claude — thisiskp_ · 2026-10-20
- Notion AI ships major upgrade: smarter Agents, Auto 90% cheaper, BYOK support — ivanhzhao · 2026-10-20
- iPhone Action Button + Grok voice + HeyGen MCP turns a sentence into an avatar video — HeyGen · 2026-10-20
- Vibe-coded game League of Spawn adds Arenas: 15-minute MOBA team fights — majidmanzarpour · 2026-10-20
- esp32sim-mcp: MCP server lets AI agents build and test ESP32 firmware in a local emulator — EmanueleStrazzullo · 2026-10-20
- GameSummoner one-line local install: prompt-to-game pipeline runs fully local on a 5090 — demonicpigg · 2026-10-20