Apple's SCLATE Gives Continual-Learning Agents One Shared Scheduler
Apple researchers have introduced SCLATE, an execution substrate for training and evaluating continual-learning agents, which compresses month-long scenarios into hours using a hybrid simulated clock. According to the paper, benchmarks and unmodified agents each add their events to one open scheduler through an adapter.
- SCLATE lets benchmarks and unmodified agents add events to one open scheduler through an adapter.
- A hybrid simulated clock skips idle gaps and compresses a month-long scenario into hours.
- Seven benchmarks were ported and ten harness and memory configurations were compared on ten models.
- An added memory system did not reliably beat the harness's native memory.
- Post-trained Qwen3.5-4B read 6.8× fewer file lines and gained 16.7 points on SWE-bench Verified.
The problem: agents that live across many sessions
Continual-learning agents, as the paper describes them, are systems of models, harnesses and memory that operate over long, multi-session horizons. Testing or training such systems means interleaving benchmark tasks with events that come from the agent itself, such as session stop and start, crons and memory consolidation.
The authors say existing benchmarks and training frameworks schedule only the benchmark's own events. As a result, each pairing of a benchmark and an agent needs its own custom scheduling loop.
How SCLATE works
SCLATE is built so that benchmarks and unmodified agents each register their events with one open event scheduler through an adapter. A hybrid simulated clock then runs all of these events on a shared timeline. The clock flows in real time while the agent is working and skips idle gaps, which, according to the authors, compresses a month-long scenario into hours.
The system also works as a rollout engine. It runs any agent's harness and memory unmodified and records the tokens and log probabilities of every model call through an in-container proxy. That record is what makes it usable for post-training as well as evaluation.
What the comparison found
The team ported seven benchmarks to SCLATE and compared ten unmodified harness and memory configurations head to head on ten models. Two findings stand out, according to the paper:
- An added memory system does not reliably beat the harness's native memory.
- Models differ widely in how they use the same harness and memory.
Post-training Qwen3.5-4B
The researchers also post-trained Qwen3.5-4B through unmodified harnesses and memory systems. The paper reports that the model learned to use both. It read 6.8× fewer file lines and reached a 16.7-point higher SWE-bench Verified pass rate. It also wrote richer memory records, and its held-out MetaClaw accuracy rose by up to 11.8 points.
What it means
For teams building agents with persistent memory, the paper points to a single shared way to schedule benchmark and agent events instead of writing a custom loop for each pairing. The comparison results also suggest that adding a memory system should be tested rather than assumed to help, since it did not reliably outperform a harness's native memory. The Qwen3.5-4B experiment indicates that the same substrate can feed training as well as evaluation, using agents as they are deployed. The source does not say whether or when SCLATE will be released.
