Apple Researchers Introduce SCLATE for Continual-Learning Agent Training and Evaluation

Apple researchers introduced SCLATE, an execution substrate designed to train and evaluate continual-learning agents that operate over long, multi-session horizons. Existing benchmarks often fail to interleave agent-side events like session stops, crons, and memory consolidation, forcing developers to build custom scheduling loops. SCLATE solves this by allowing benchmarks and unmodified agents to add events to a single open scheduler via adapters.

Efficient Simulation and Rollout

The system uses a hybrid simulated clock that runs events on a shared timeline, flowing in real time while the agent works but skipping idle gaps. This approach compresses month-long scenarios into hours. SCLATE also functions as a rollout engine, running any agent’s harness and memory unmodified while recording tokens and log probabilities for every model call through an in-container proxy.

Benchmarking and Post-Training Results

The team ported seven benchmarks to SCLATE and compared ten unmodified harness and memory configurations across ten models. Key findings include:

• Added memory systems do not reliably outperform a harness’s native memory.
• Models vary significantly in how they utilize the same harness and memory.

Researchers post-trained Qwen3.5-4B using unmodified harnesses and memory systems. The model learned to use both effectively, resulting in:

• 6.8× fewer file lines read.
• 16.7-point higher pass rate on SWE-bench Verified.
• Up to 11.8-point rise in held-out MetaClaw accuracy.