feat: integrate MemPalace as persistent agent memory backend

[ductrantrong]

Proposal

Integrate MemPalace as a persistent memory backend for GoClaw agents — giving them cross-session recall with the highest retrieval accuracy ever benchmarked.

Why MemPalace

MemPalace holds the highest LongMemEval score ever published, free or paid:

Metric	Score
LongMemEval R@5 (zero API calls)	96.6%
LongMemEval R@5 (with Haiku rerank)	100%
Retrieval boost from palace structure	+34%
Cost	$0 — local only, no cloud, no subscription

Benchmarks are reproducible: benchmarks/

Why This Fits GoClaw

1. MCP-native — MemPalace ships an MCP server (python -m mempalace.mcp_server) exposing 19 tools. GoClaw already supports MCP (stdio/SSE/streamable-http), so the integration path is straightforward: register MemPalace as an MCP tool provider per agent or per tenant.

2. Complements existing memory — GoClaw has BM25 + pgvector hybrid search for skills and a knowledge graph for structured facts. MemPalace adds a different layer: lossless conversational memory organized into a palace structure (wings → halls → rooms). Agents could recall why a decision was made months ago, not just what was decided.

3. AAAK compression — MemPalace's AAAK dialect compresses months of context into ~170 tokens for agent wake-up. For GoClaw's multi-tenant setup, this means each agent can load its full history cheaply at session start without blowing token budgets.

4. Multi-tenant alignment — MemPalace stores everything locally per project/person. This maps naturally to GoClaw's per-user PostgreSQL workspaces — each tenant gets their own palace.

5. Local-first — No external API calls needed. Runs on ChromaDB locally. Fits GoClaw's "$5 VPS" deployment target.

Suggested Implementation

Phase 1: MCP Tool Provider

• Register MemPalace MCP server as an optional tool provider in agent config
• Expose mempalace_search, mempalace_mine, mempalace_status etc. to agents
• Per-tenant palace paths mapped to GoClaw workspaces

Phase 2: Native Integration

• Auto-mine agent conversations into MemPalace after each session
• AAAK wake-up context injected into agent system prompts on session start
• Palace management via GoClaw admin dashboard

Phase 3: Deep Integration

• Replace or augment pgvector memory with MemPalace's palace-structured retrieval for conversation history
• Team-level palaces for shared agent memory across agent teams
• Heartbeat agents that periodically mine and organize palace data

References

• MemPalace repo: https://github.com/milla-jovovich/mempalace
• GoClaw MCP docs: https://docs.goclaw.sh (MCP integration section)
• LongMemEval benchmark paper context in MemPalace's benchmarks/BENCHMARKS.md