SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation
Apple Machine Learning Research published a paper titled "SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation." The authors are Youngmok Jung, Sirajul Salekin, Henry Tran, Javier Movellan, Zhao Huang, and Manjot Bilkhu. It is listed under the research areas Methods and Algorithms and Tools, Platforms, Frameworks.
The paper states that continual-learning agents are systems of models, harnesses, and memory operating over long multi-session horizons, and that evaluating and training them requires interleaving tasks with agent-side events such as session stop and start, crons, and memory consolidation. It says existing benchmarks and training frameworks schedule only the benchmark's own events, leaving each benchmark and agent pair to build a custom scheduling loop.
SCLATE is described as an execution substrate where benchmarks and unmodified agents each add their events to one open event scheduler through an adapter. A hybrid simulated clock runs these events on a shared timeline, flowing in real time while the agent works and skipping idle gaps, which the paper says compresses a month-long scenario into hours. SCLATE also serves as a rollout engine that runs any agent's harness and memory unmodified, recording the tokens and log probabilities of every model call through an in-container proxy.
The paper reports porting seven benchmarks to SCLATE and comparing ten unmodified harness and memory configurations head to head on ten models. That comparison shows that an added memory system does not reliably beat the harness's native memory, and that models differ widely in how they use the same harness and memory. The authors then post-train Qwen3.5-4B through unmodified harnesses and memory systems, reporting that the model learns to use both, reading 6.8x fewer file lines with a 16.7-point higher SWE-bench Verified pass rate, and writing richer memory records, while its held-out MetaClaw accuracy rises by up to 11.8 points.
The paper was published in September 2026.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Continual-learning agents are systems of models, harnesses, and memory operating over long multi-session horizons. Evaluating and training them requires interleaving tasks with agent-side events such as session stop and start, crons, and memory consolidation. Yet existing benchmarks and training frameworks schedule only the benchmark’s own events, leaving each benchmark and agent pair to build a custom scheduling loop. We present SCLATE, an execution substrate where benchmarks and unmodified agents each add their e