Create Batch

Import up to 500 memories in one call.

Every row is checked as a single write is (its content, metadata, timestamp, user_id, agent_id and run_id) before anything is queued: a bad row is refused with its index in the message, and nothing is imported. Then the job is queued and the call answers 202 with its job_ id. The worker writes the rows, which get their facts as any write does; GET /v1/batch/{job_id} says how far it is. If the key is revoked, expires or is moved to another project before the worker is done, the rows not yet imported are not: each is a failed row with its reason in errors, and the job completes.