update

Usage

machine-memory update <id> <content> (--local|--remote)
machine-memory update <id,id,...> --from-file <path> (--local|--remote)
machine-memory update --match "unique topic" <content> (--local|--remote)

Provide either one or more IDs or --match <query>, then provide inline content or --from-file <path>. Updating multiple IDs returns the updated rows, missing IDs, and a count.

Metadata

--tags, --context, --type, --certainty, --updated-by, --refs, --evidence, and --validation-gaps replace the corresponding metadata; omitted flags keep the stored values. --evidence none and --validation-gaps none clear them. --expires-after-days <n|null> sets or clears expiry.

Setting --certainty verified, or clearing the evidence of a verified memory, requires at least one evidence kind (see add). Memories saved before evidence existed keep their certainty with no evidence; content-only updates to them still work, and doctor lists them for review. The content update increments update_count and refreshes updated_at.

--match must resolve to an active memory and cannot be combined with more than one positional content argument.

Reconciliation

When the content changes, the response reports:

• dropped: identifiers and sentences the new content no longer carries, as in diff. --dry-run reports them per target before anything is written.
• review_candidates: up to five other active memories that may rely on the old premise, with match_strength. A strong candidate names the same path or identifier. Update or deprecate each one, or confirm it still holds. Status memories do not trigger this check.

--json-min and --quiet keep dropped and list review_candidate_ids instead of summaries; batch updates group both per ID under reconciliation. add --upsert-match reports the same fields when it updates a strong match.

Updates save the previous row automatically. Use history and restore to review or recover an accidental overwrite without a prior read.

Embedding budget

Like add, update validates the composed embedding text before writing. If the updated memory (new content plus retained or replaced metadata) would exceed the 512-token limit — or the conservative byte+2 estimate — the update fails with the exact byte or token deficit and no memory is changed. Preflight with --dry-run to print a per-target size report without writing; it exits 1 when any target would be over budget:

machine-memory update 42 "replacement content" --dry-run --remote
# {"command":"update","dry_run":true,"count":1,
#  "targets":[{"id":42,"size":{...}}]}

Output flags: --brief prints {id, status} (plus tokens with --token-report); --json-min/--quiet print {id} for single targets and compact {updated_ids, not_found, count} payloads for batches.

Pass --token-report to include the breakdown on a successful update as tokens (for a single target) or as a tokens array of {id, ...breakdown} entries when updating multiple IDs.