Meta paper: routing agent memory by task type beats shared history, cuts API cost 65-72%
rohanpaul_ai · x · 2026-11-03
A new FAIR at Meta paper introduces Routed Continual Learning (Routed CL): instead of one shared history, an LLM-based router decides before each run whether to reuse an existing memory cluster or start a new one, with each cluster using one of 5 memory systems (raw trajectories, distilled lessons, etc.).
Evaluated on 398 mixed problems (CL-Bench + τ³-Bench banking tasks) across 4 models:
- A single global history barely beat no memory (mean gain 0.015 out of 1)
- Routed CL outperformed global history by 15-30% of the maximum attainable normalized reward
- Estimated API cost dropped 65-72%
Takeaway: agents handling mixed task types should split memory into separate pools by problem type.
Related event: Meta's Routed Continual Learning Cuts Agent Memory API Costs by Up to 72%(2 posts)→
More from coding & agent
- DeepSeek's super-agent explained: one router, many parallel coding workers — zainhas · 2026-11-03
- WonderSearch lets Claude Code, Codex and Cursor share one brain—no vector DB needed — Scobleizer · 2026-11-03
- Xbow breaks down the $40K GPT-5.6-sol API bill behind its full-chain Chrome exploit — moyix · 2026-11-03
- Phone Harness MCP lets any MCP agent control real cloud iPhones and Androids — Scobleizer · 2026-11-03
- LangChain on late-2026 context engineering: browser use, long-term goals, async work — hwchase17 · 2026-11-03
- MotherDuck launches Flight Explainer: live charts for agent-built Python pipelines — josh_wills · 2026-11-03