Meta Paper: Training Agents to Decide When to Use Memory via RL
rohanpaul_ai · x · 2026-08-24
Meta introduces EvoHarness-RL, a novel method addressing the issue of long-horizon agents forgetting state due to task length.
Key Insight: Simply adding memory and tools is insufficient; the critical breakthrough is training the agent when it is worth using external state. EvoHarness-RL uses Reinforcement Learning to teach the policy model to balance between "calling external state (Belief, Progress, Experience)" and "saving interaction steps."
Performance:
- On ALFWorld tasks, Qwen3-8B with ReAct scored only 47.9%. With EvoHarness-RL, it surged to 96.9% (seen tasks) and 86.6% (unseen tasks).
- During training, as success rates rose, the frequency of external state calls dropped to approximately 1 per episode. This indicates routine behavior is internalized into the model weights, leaving external memory access for critical moments.
Takeaway: Agent builders should train policies to decide when to consult and update memory rather than hard-coding constant access.
Related event: Meta's EvoHarness-RL Trains Agents to Decide When to Use Memory(2 posts)→
More from coding & agent
- Vercel Goes Full Backend: Python GA, Containers, and Daemons — cramforce · 2026-10-16
- WorldMemArena Accepted at NeurIPS: Memory Storage ≠ Memory Use — xwang_lk · 2026-10-16
- Vercel Ship Demos Container Deploys for Go, Ruby, Java and More — edwin · 2026-10-16
- TRL adds day-0 support for fine-tuning Muse Spark Open's 816B MoE with LoRA on 64 H100s — QGallouedec · 2026-10-16
- Databricks engineers land impactful PRs on day 2 with AI dev tools, vs week 3 before — Yuchenj_UW · 2026-10-16
- Perplexity open-sources SPACE, the sandbox platform powering its Computer agents — AravSrinivas · 2026-10-16