openclaw - 💡(How to fix) Fix Proposal: Palace-style structured memory — integrate hierarchical memory architecture (inspired by MemPalace) [1 participants]

Motivation

Current OpenClaw memory is flat: MEMORY.md (curated long-term) + memory/YYYY-MM-DD.md (daily logs). This works, but has fundamental limitations:

1. Flat files hit context limits — MEMORY.md grows until it gets truncated. The 20KB soft cap is a band-aid, not a solution.
2. No structure = poor retrieval — semantic search over a flat file finds matches but can't leverage organizational hierarchy.
3. Claude's own context management (evident in Claude Code's architecture leaks) shows that structured, categorized memory significantly outperforms flat storage.

Proposal: Palace Memory Architecture

Inspired by MemPalace — which benchmarks 96.6% recall on LongMemEval with a hierarchical "Palace" structure.

Core Design

memory/
  INDEX.md              ← Short curated index (search entry point)
  palace/
    personal/            ← Wing: people, relationships
      preferences/     ← Hall: user preferences, habits
      decisions/        ← Hall: key decisions made
    projects/           ← Wing: work, tasks, goals
      milestones/       ← Hall: achievements, progress
      problems/         ← Hall: technical issues, solutions
    knowledge/          ← Wing: learned facts, concepts
      concepts/         ← Hall: explanations, frameworks
      references/       ← Hall: links, resources

Key insight from MemPalace: the structure itself improves retrieval by 34% — before any semantic search.

OpenClaw-Specific Advantages

1. Pre-compaction flush already exists — compaction.memoryFlush already auto-triggers before context compaction. Point this at the Palace structure instead of flat files.
2. Built-in vector search — OpenClaw's memory_search supports hybrid BM25 + vector search + MMR. Apply it to Palace subdirectories for domain-specific retrieval.
3. No new dependencies — MemPalace's gains come from structure, not new libraries. The Palace structure is just organized Markdown files.
4. MMCP integration — MemPalace has an MCP server; OpenClaw v4.x supports MCP. A native Palace MCP provider could pull in MemPalace's research directly.

Compatibility

• Backwards compatible: existing MEMORY.md and memory/*.md remain valid
• Gradual migration: new memory goes to Palace structure, old files remain searchable
• memory_search works: vector index can span directory trees

Comparison

Implementation Ideas

1. Config option agents.defaults.memory.layout: "flat" | "palace" — flat is default, palace opt-in
2. Auto-classify on write — when agent writes to memory/YYYY-MM-DD.md, also classify into Palace subdirectories
3. Palace-aware memory_search — prioritize results from relevant Wing/Hall based on query terms
4. MMCP integration — OpenClaw v4.x supports MCP; connect to MemPalace MCP server for native Palace support

References

• MemPalace: https://github.com/milla-jovovich/mempalace
• Benchmark: 96.6% LongMemEval R@5, +34% retrieval boost from structure alone
• AAAK dialect: 30x lossless compression for AI context (relevant if Palace content grows large)

Inspired by hands-on use of OpenClaw alongside MemPalace. Happy to help test or prototype.