Copy-on-Write Isolation

When steps run in parallel, each one needs its own view of memory so siblings don’t clobber each other. CARL does this with a copy-on-write (COW) store instead of deep-copying the whole memory for every step.

What you need to know

• Each parallel step sees the memory as it was at the start of its batch.
• A step’s writes are not visible to its parallel siblings — only to subsequent batches.
• Tools shared across parallel steps must be stateless (they’re shared via a shallow copy).

This is automatic — there’s nothing to configure. It just makes parallel execution correct and cheaper than full deep copies.

Profiling

Per-step memory and history sizes are recorded in each StepExecutionResult.profiling (keys memory_bytes_after, history_bytes_after, history_bytes_added) and aggregated by result.get_profiling_summary(), so you can see how memory grows across a run.

See also

• Async execution — the parallelism model.
• Memory overview