Problem

OpenClaw's memory system uses a two-layer architecture:

• memory/YYYY-MM-DD.md — daily append-only logs (short-term)
• MEMORY.md — curated long-term memory

However, the transfer from short-term to long-term memory is entirely manual. The agent must be explicitly reminded to consolidate memories. There is no automated mechanism for identifying high-value information in daily logs and promoting it to long-term memory.

Proposal: Automated Memory Consolidation

Inspired by the hippocampal-cortical memory consolidation process in neuroscience:

Phase 1: Tagging (during session)

Automatically tag high-signal entries in daily logs based on:

• Explicit user requests ('remember this')
• Repeated references to the same entity/topic
• Decision points and their outcomes
• Emotional markers (user frustration, satisfaction)

Phase 2: Consolidation (periodic background job)

A cron-based consolidation job that:

1. Scans recent daily logs for tagged entries
2. Deduplicates against existing MEMORY.md content
3. Summarizes and structures new entries
4. Appends to MEMORY.md with source attribution

Phase 3: Forgetting (memory hygiene)

Implement an Ebbinghaus-inspired decay:

• Entries not referenced in N days get confidence-downgraded
• Low-confidence entries are archived (not deleted)
• Frequently referenced entries get confidence-upgraded

Related Work

• docs/experiments/research/memory.md — Memory v2 research notes already propose a bank/ architecture with typed memory pages
• memory-core plugin — current search/get tools
• compaction.memoryFlush — existing pre-compaction memory write

Impact

• Reduces manual memory management burden on agents
• Improves recall quality over time
• Aligns with the existing Memory v2 research direction