uTTT: a fast-weight memory shared across all layers enables test-time learning for lifelong agents
JunjieHu12 · x · 2026-10-08
Researchers including Junjie Hu unveiled uTTT (Universal Test-Time Training), built on a simple idea: model memory doesn't need to belong to individual layers. Instead, one fast-weight memory is shared by every layer of the network, using the same state and compute.
Key points:
- A single shared memory module is reused by all layers, replacing per-layer memories
- Improves long-context retrieval and view synthesis
- Enables rapid learning at test time
- The authors frame it as a step toward lifelong agents that can continually learn from experience after deployment, addressing the challenge of tracking ever-longer agent trajectories
Paper, code, and project page are available.
More from coding & agent
- Jeffrey Emanuel's "say no to process" agent skill kills Codex ceremony output — used hundreds of times a day — doodlestein · 2026-10-08
- a16z backs Preference Model, which open-sources Karotte RL environment framework battle-tested by 1M+ evals — a16z · 2026-10-08
- Every's agent skims meeting notes and only pings you when your name comes up — here's the 4-step setup — every · 2026-10-08
- Exa's setup page swaps dev docs for a copy-paste prompt your coding agent runs — josh_bickett · 2026-10-08
- Haiku 5.5 targets high-volume tasks, works as a coding subagent with Opus/Sonnet — claudeai · 2026-10-08
- Non-coder runs his entire business on an army of Claude Opus 5.5 agents — EXM7777 · 2026-10-08