Neutral memory interchange and Mem0 migration

AtMem's line-delimited JSON archive begins with an atmem-memory-archive-v1 manifest followed by individually digested neutral records. Each record carries an explicit tenant, subject, workspace and agent scope, lifecycle state, source metadata, evidence and supported history.

Mem0 mapping preserves source IDs, content, metadata and source timestamps. Unknown vendor fields are counted in unsupported_fields; vectors, graphs, credentials and undocumented fields are never trusted as canonical data. Missing scope is an error, not a broad default.

Imports always start with deterministic dry_run(). Conflicts and sensitive records require explicit review. Commits use atomic batches, stable source IDs, canonical generation changes and durable checkpoints. An interruption resumes after the last acknowledged batch, while an exact replay returns the same run receipt. Rollback tombstones only record IDs created by that run.

Exports reapply current authorization and lifecycle at execution time. Their receipt binds the archive bytes and record-set digest. Back up canonical state before a large migration; rollback does not restore foreign fields listed as unsupported in the mapping report.