AI compression ("summarize") — governance + operator contract

This doc defines how we use LLM-assisted compression in openclaw-mem without turning it into a silent-footgun.

What AI compression is

Compression turns high-volume, low-signal logs into a smaller set of derived notes.

It is not a replacement for retrieval. It is a way to keep the operator surface readable.

Decision (default posture)

• Compression output is a derived artifact by default.
• We do not auto-promote compressed text into durable memory.
• We do not overwrite operator-authored files.

Rationale: a bad summary is worse than no summary, because it creates false confidence.

Required guardrails

1) Deterministic input selection

• Select a bounded input window (e.g., one day, or N records).
• Prefer importance/trust-aware selection (must > nice > unknown, cap the rest).

2) Hard caps (cost + length)

• Always set a hard output limit (e.g., max_tokens <= 700).
• Limit frequency (daily or less) unless explicitly justified.

3) Receipts (must be debuggable)

Every compression run must record: - time window / input selection rules - model + max_tokens - output length - a pointer to the derived artifact path

4) Fail-open

If compression fails: - ingest/harvest/triage must still work - the system should remain operable (no broken pipeline)

5) Rollback

Rollback is simple: - ignore/delete the derived artifact - re-run compression from raw sources

Implementation status

• openclaw-mem summarize exists (CLI).
• The governance contract above is the acceptance bar for making it part of daily ops.

Future: promotion (explicit, reviewed)

If we later decide to promote compressed notes into durable memory, it must be: - explicit (operator-reviewed) - reversible (archive-first) - tracked (receipts + provenance)