openclaw - 💡(How to fix) Fix memory: add lifecycle-aware LLM curation for MEMORY.md

Problem

Root MEMORY.md is intended to be curated long-term memory, but current automated promotion only selects and formats short-term recall candidates. Even with compact snippets, it does not manage memory lifecycle:

• some entries are durable anchors and should be protected from automatic deletion or rewriting;
• some entries are evolving facts and should be refreshed when daily evidence changes;
• some entries become stale and should be downgraded, archived, or removed from root memory;
• provenance should remain visible so humans and future agents can inspect why an entry exists.

This makes MEMORY.md vulnerable to slow drift: raw or outdated facts can accumulate, while newer daily evidence does not naturally replace older conclusions.

Proposed Direction

Add an explicit long-term memory curator stage, likely LLM-assisted, separate from deterministic short-term promotion.

The curator should classify root memory entries by lifecycle policy, for example:

• locked: stable identity, safety, privacy, or owner preference anchors; never auto-delete or rewrite without explicit approval.
• stable: durable conclusions that can be retained unless contradicted by strong newer evidence.
• evolving: operational/project facts that should be periodically checked against recent daily notes and refreshed when the current state changes.
• stale: entries no longer supported by recent evidence or superseded by newer notes; candidates for archival/removal.

Requirements

• Keep daily notes as the raw append-only history.
• Keep MEMORY.md compact and curated, not a dump of recalled daily chunks.
• Preserve provenance metadata such as score, recalls, avg, and source for promoted entries.
• Support a configurable input budget for daily evidence considered by the curator.
• Produce a reviewable patch/diff rather than silently rewriting root memory.
• Respect locked sections/anchors and require explicit approval before modifying them.
• Prefer deterministic safeguards around the LLM output: schema/output contract, max tokens, validation, and rollback-safe writes.

Relationship To #87654

PR #87654 is a narrow mitigation: it bounds promoted snippets and adds configurable dreaming.phases.deep.maxPromotedSnippetTokens while preserving provenance.

This issue tracks the deeper fix: semantic/lifecycle curation of long-term memory.