Skip to main content
<MemoryTrellis> is <Trellis> wrapped in <Memory>. Initial and continuation authors, validation, workers, settlement, and the final task all inherit the same normalized memory policy.
The component accepts every TrellisProps field plus one required memory field containing MemoryProps without children:

​
Runtime guarantees

Each generated task receives the policy in TaskDescriptor.memoryConfig, so scorers and eval reports can segment results by bank, tags, recall mode, retention mode, or tool access. Give experiment arms stable tags when they must remain isolated. The standard memory runtime supplies the execution guarantees:
  • recalled context and primers share the maxTokens bound
  • recall failures run the original task prompt instead of failing the task
  • one task execution keeps its recalled snapshot frozen across retries
  • retain="on-complete" writes asynchronously after successful completion
See <Memory> for bank filtering, backend selection, and retention details. Trellis still requires the run’s pinned concurrency to match maxConcurrency.

​
Experiment arms

Keep all non-memory Trellis props identical when comparing arms.

​
Baseline

​
Recall only

​
Recall and learn