MemPalace — Local-First AI Memory Palace

Give your AI a memory. Mine projects and conversations into a searchable palace. No API key required. Exposed via MCP (Model Context Protocol) and CLI.

Version: 3.5.0 | License: MIT

Forked from MemPalace/mempalace — đang được kiểm tra, nâng cấp và vận hành bởi @SlncTrZ (Trương Công Định)

---

⚠️ Trạng thái dự án

Dự án này đang trong quá trình kiểm tra, sửa lỗi và nâng cấp. Một số thành phần có thể chưa ổn định hoặc đang được tái cấu trúc.

---

🚀 Nâng cấp từ bản gốc

Khoản mục	Bản gốc	Bản này
MCP Tools	~10 tools	30 tools — Palace R/W, Knowledge Graph, Navigation, Agent Diary, Hooks
Storage Backend	Qdrant-only	3 backends — Qdrant (default), PgVector, SQLite Exact
Embedding	Ollama-dependent	ONNX local (zero-dep) + Ollama (optional)
Knowledge Graph	❌ Không có	Temporal entity-relationship graph (SQLite, valid_from/valid_to, invalidation)
Palace Navigation	❌ Không có	Room graph, BFS traversal, cross-wing tunnels, halls
Agent Diary	❌ Không có	AAAK-compressed diary per agent, read/write tools
i18n	❌ Không có	15 ngôn ngữ (EN, VI, JA, KO, ZH, FR, DE, ES, RU, IT, PT-BR, HI, ID, BE, ZH-TW)
Sources Framework	❌ Không có	RFC 002 — extensible source adapters for mining
Auto-Save Hooks	❌ Không có	Stop hook (15 msg), PreCompact hook
AAAK Dialect	❌ Không có	Entity codes, emotion markers, 30× lossless compression
Entity System	❌ Không có	Auto-detect people/projects, entity code registry
CLI	5 subcommands	12+ subcommands (init, mine, search, split, wake-up, compress, status, repair, mcp, hook, instructions)
Modules	8 modules	32 modules — thêm repair, backups, sweeper, migrate, daemon, dynamics, exporter, fact_checker, format_miner, corpus_origin, dedup, hallways, ids, sync, wal, closet_llm, convo_scanner, diary_ingest, spellcheck, sources
Tests	~20 (lỗi import)	128 tests pass — protocol, backend registry, Qdrant, palace, config, AAAK
HTTP Server	❌ Không có	SSE transport + Qdrant query API + landing page
Security	API key trong code	Pre-commit hook, env vars, .gitignore, không hardcoded IP
CLI Entry Point	Không rõ ràng	mempalace + mempalace-mcp (pyproject.toml scripts)

---

Architecture

┌──────────────────────────────────────────────────┐
│ MCP Client (Claude Code / Cline) │
│ ↓ stdio JSON-RPC │
├──────────────────────────────────────────────────┤
│ mcp_server.py — 30 MCP tools │
├──────────────────────────────────────────────────┤
│ CLI (mempalace) ←→ Palace (Qdrant) │
│ ↕ │
│ Knowledge Graph (SQLite, temporal) │
│ ↕ │
│ Palace Graph (room navigation, tunnels) │
├──────────────────────────────────────────────────┤
│ Backends: Qdrant | PgVector | SQLite Exact │
│ Embeddings: ONNX (local) | Ollama | external │
└──────────────────────────────────────────────────┘

Palace Structure

Wings (projects/people)
└── Rooms (topics)
└── Closets (summaries)
└── Drawers (verbatim memories)
Halls → connect rooms within a wing
Tunnels → connect rooms across wings

MCP Tools — 30 tools

Palace (Read)

Tool	Description
mempalace_status	Palace overview — total drawers, wing & room counts
mempalace_list_wings	List all wings with drawer counts
mempalace_list_rooms	List rooms within a wing (or all)
mempalace_get_taxonomy	Full taxonomy: wing → room → drawer count
mempalace_search	Semantic search — query, wing/room filter, distance threshold
mempalace_check_duplicate	Check if content already exists before filing
mempalace_get_drawer	Fetch a single drawer by ID (full content + metadata)
mempalace_list_drawers	List drawers with pagination, wing/room filter
mempalace_get_aaak_spec	Get the AAAK specification — ⚠️ lossy summary, not lossless compression

Palace (Write)

Tool	Description
mempalace_add_drawer	File verbatim content into a wing/room
mempalace_update_drawer	Update an existing drawer's content / metadata
mempalace_delete_drawer	Delete a drawer by ID (irreversible)
mempalace_sync	Prune drawers from deleted/moved source files

Knowledge Graph (Temporal Facts)

Tool	Description
mempalace_kg_query	Query entity relationships (outgoing/incoming/both) with temporal filter
mempalace_kg_add	Add a fact: subject → predicate → object (optional valid_from/valid_to)
mempalace_kg_invalidate	Mark a fact as no longer true
mempalace_kg_timeline	Chronological timeline of facts for an entity (or all)
mempalace_kg_stats	Knowledge graph overview: entities, triples, current vs expired

Navigation (Palace Graph)

Tool	Description
mempalace_traverse	Walk the palace graph from a room (BFS, max_hops)
mempalace_find_tunnels	Find rooms bridging two wings
mempalace_create_tunnel	Create a cross-wing tunnel
mempalace_list_tunnels	List all explicit tunnels (optional wing filter)
mempalace_delete_tunnel	Delete a tunnel by ID
mempalace_follow_tunnels	Follow tunnels from a room to connected wings
mempalace_graph_stats	Graph overview: rooms, tunnels, edges between wings

Agent Diary

Tool	Description
mempalace_diary_write	Write to agent diary in AAAK format
mempalace_diary_read	Read recent diary entries

Settings & Hooks

Tool	Description
mempalace_hook_settings	Get/set hook behavior (silent_save, desktop_toast)
mempalace_memories_filed_away	Check if a recent checkpoint was saved
mempalace_reconnect	Force reconnect after external writes

CLI Commands

mempalace init <dir> Initialize a new palace
mempalace mine <dir> Mine a project (default mode)
mempalace mine <dir> --mode convos Mine conversation exports
mempalace search "query" Search your memories
mempalace split <dir> Split large transcript files
mempalace wake-up Load palace into context
mempalace compress Compress palace storage
mempalace status Show palace status
mempalace repair Qdrant snapshot backup & recovery
mempalace check-qdrant Check Qdrant backend health
mempalace mcp Show MCP setup command
mempalace hook run Run hook logic (for harness)
mempalace instructions <name> Output skill instructions

Setup

pip install -e ".[dev]"# Or: pip install mempalace

Environment Variables

Variable	Default	Description
MEMPALACE_HOME	~/.mempalace	Palace data directory
MEMPALACE_EMBEDDING_DEVICE	auto	Embedding device: cpu, cuda, dml, coreml
PALACE_BACKEND	qdrant	Storage backend: qdrant, pgvector, sqlite_exact
PALACE_PATH	(auto)	Explicit palace path
QDRANT_URL	http://localhost:6333	Qdrant server URL (qdrant backend)
OLLAMA_URL	http://localhost:11434	Ollama server URL (optional)

Backends

Backend	Purpose
Qdrant (default)	Vector DB — local Qdrant or remote server
PgVector	PostgreSQL vector extension — for existing Postgres deployments
PgVector	PostgreSQL vector extension — for existing Postgres deployments
SQLite Exact	Exact (brute-force) search — debug, tiny palaces, CI

Embedding Options

• Local ONNX model (default) — zero-dependency, no API key
• Ollama — nomic-embed-text:latest (768d) or other models
• External embedding API (custom adapters)

Key Modules

Module	Description
cli.py	CLI entry point — 12+ subcommands
mcp_server.py	MCP server handler (imports from mcp_tools.py)
mcp_tools.py	30 MCP tool definitions + handlers (tách từ monolith)
miner.py	Project file ingest — chunks by paragraph
convo_miner.py	Conversation ingest — exchange pairs, room detection
searcher.py	Semantic search with filters
layers.py	4-layer memory stack (L0 identity → L3 deep search)
dialect.py	AAAK compression — entity codes, 30× lossless ratio
knowledge_graph.py	Temporal entity-relationship graph (SQLite)
palace_graph.py	Room-based navigation graph (BFS, tunnels)
entity_registry.py	Entity code registry — AAAK codes, ambiguous names
entity_detector.py	Auto-detect people/projects from file content
normalize.py	Convert 5 chat formats to standard transcript
onboarding.py	Guided first-run setup
backups.py	Palace backup & restore
repair.py	Rebuild corrupted vector indices
sweeper.py	Clean up stale/expired drawers
migrate.py	Schema migration across versions

Test

python -m pytest tests/ -v

Scripts

Script	Purpose
scripts/pre_commit_check.py	Pre-commit API key scanner

Security

See SECURITY.md for API key protection and pre-commit hook setup.

⚠️ Hạn chế hiện tại

AAAK Dialect — Lossy Summary, không phải nén lossless

AAAK (dialect.py) là định dạng tóm tắt có cấu trúc (structured summary), không phải nén lossless:

• Nén (Compression): Xử lý cục bộ bằng Python (dialect.py) — dùng regex, keyword mapping, emotion dictionary, entity extraction. Không gọi LLM.
• Giải nén (Decompression): Format được thiết kế để LLM đọc trực tiếp — không có decoder. LLM suy diễn lại ngữ cảnh từ các token còn lại, dẫn đến hallucination.
• Dữ liệu gốc: Verbatim content được lưu riêng trong drawers (ChromaDB/Qdrant). AAAK chỉ là summary layer để định hướng tra cứu, không thay thế dữ liệu gốc.
• Xem docstring đầu file mempalace/dialect.py để biết chi tiết.

Các hạn chế khác

Hạn chế	Mô tả
CLI monolith	cli.py ~1.8k lines — nhiều subcommand trong 1 file
Thiếu type hints	Hầu hết các module chưa có type annotations đầy đủ
Test coverage	Chỉ ~9% — cần integration tests với Qdrant thật
Documentation	Chưa có API docs tự động (Sphinx/MkDocs)
Chưa publish PyPI	Chưa có release trên PyPI

License

MIT