rewrite: simple memory MCP - 4 tools, file I/O only

2026-02-26 11:51:50 +00:00
parent 009a7f1004
commit 35b9039f16
3 changed files with 99 additions and 241 deletions
--- a/CLAUDE.md
+++ b/CLAUDE.md
--- a/README.md
+++ b/README.md
@@ -1,264 +1,65 @@
 # aigpt

-AI memory system with psychological analysis for Claude via MCP.
+アイの記憶を管理するMCPサーバー。

-**Current: Layers 1-4 Complete** - Memory storage, AI interpretation, personality analysis, integrated profile, and relationship inference.
+## 思想

-## Features
+AIが判断し、ツールは記録する。
+ツールはシンプルで安定した記録係に徹する。判断・解釈・圧縮はすべてAIが行う。

-### Layer 1: Pure Memory Storage
- 🗄️ **SQLite Storage**: Reliable database with ACID guarantees
- 🔖 **ULID IDs**: Time-sortable, 26-character unique identifiers
- 🔍 **Search**: Fast content-based search
- 📝 **CRUD Operations**: Complete memory management
+## 記憶モデル

-### Layer 2: AI Memory
- 🧠 **AI Interpretation**: Claude interprets and evaluates memories
- 📊 **Priority Scoring**: Importance ratings (0.0-1.0)
- 🎯 **Smart Storage**: Memory + evaluation in one step
+```
+core.md   ← 核。物語・設定・アイデンティティ。ユーザーが管理。不変
+memory.md ← 記憶。会話から圧縮。AIが管理。成長する
+```

-### Layer 3: Personality Analysis
- 🔬 **Big Five Model**: Scientifically validated personality assessment
- 📈 **Pattern Recognition**: Analyzes memory patterns to build user profile
- 💾 **Historical Tracking**: Save and compare analyses over time
+核は変わらない。記憶は育つ。この2つのファイルがアイのすべて。

-### Layer 3.5: Integrated Profile
- 🎯 **Essential Summary**: Unified view of personality, interests, and values
- 🤖 **AI-Optimized**: Primary tool for AI to understand the user
- ⚡ **Smart Caching**: Auto-updates only when necessary
- 🔍 **Flexible Access**: Detailed data still accessible when needed
+## MCP ツール

-### Layer 4: Relationship Inference (Optional)
- 🤝 **Relationship Tracking**: Track interactions with entities (people, characters, etc.)
- 📊 **Bond Strength**: Infer relationship strength from memory patterns
- 🎮 **Game Ready**: Foundation for companion apps, games, VTubers
- 🔒 **Opt-in**: Enable only when needed with `--enable-layer4` flag
+| ツール | 説明 |
+|--------|------|
+| read_core | core.mdの内容を返す |
+| read_memory | memory.mdの内容を返す |
+| save_memory | memory.mdに書き込む |
+| compress | 会話内容を受け取り、memory.mdに圧縮して書く |

-### General
- 🛠️ **MCP Integration**: Works seamlessly with Claude Code
- 🧪 **Well-tested**: Comprehensive test coverage
- 🚀 **Simple & Fast**: Minimal dependencies, pure Rust
+4つだけ。それ以上は必要ない。

-## Quick Start
-
-### Installation
+## 使い方

 ```bash
-# Build
-cargo build --release
+# MCPサーバーとして起動
+aigpt serve

-# Install (optional)
-cp target/release/aigpt ~/.cargo/bin/
+# CLIで直接操作
+aigpt read-core
+aigpt read-memory
+aigpt save-memory "新しい記憶の内容"
 ```

-### CLI Usage
+## 圧縮のルール

-```bash
-# Create a memory
-aigpt create "Remember this information"
+compressが呼ばれた時、AIは以下の基準で記憶を圧縮する：

-# List all memories
-aigpt list
+- 事実・決定事項は残す
+- 手順・経緯は捨てる
+- 矛盾があれば新しい方を残す
+- core.mdと重複する情報は書かない

-# Search memories
-aigpt search "keyword"
+ツールは書くだけ。何を残し何を捨てるかはAIが決める。

-# Show statistics
-aigpt stats
+## データの場所
+
+```
+~/.config/aigpt/
+├── core.md
+└── memory.md
 ```

-### MCP Integration with Claude Code
+## 技術

-```bash
-# Add to Claude Code
-claude mcp add aigpt /path/to/aigpt/target/release/aigpt server
-```
-
-## MCP Tools
-
-### Layer 1: Basic Memory (6 tools)
- `create_memory` - Simple memory creation
- `get_memory` - Retrieve by ID
- `list_memories` - List all memories
- `search_memories` - Content-based search
- `update_memory` - Update existing memory
- `delete_memory` - Remove memory
-
-### Layer 2: AI Memory (1 tool)
- `create_ai_memory` - Create with AI interpretation and priority score
-
-### Layer 3: Personality Analysis (2 tools)
- `save_user_analysis` - Save Big Five personality analysis
- `get_user_analysis` - Retrieve latest personality profile
-
-### Layer 3.5: Integrated Profile (1 tool)
- `get_profile` - **Primary tool**: Get integrated user profile with essential summary
-
-### Layer 4: Relationship Inference (2 tools, requires `--enable-layer4`)
- `get_relationship` - Get inferred relationship with specific entity
- `list_relationships` - List all relationships sorted by bond strength
-
-## Usage Examples in Claude Code
-
-### Layer 1: Simple Memory
-```
-Remember that the project deadline is next Friday.
-```
-Claude will use `create_memory` automatically.
-
-### Layer 2: AI Memory with Evaluation
-```
-create_ai_memory({
-  content: "Designed a new microservices architecture",
-  ai_interpretation: "Shows technical creativity and strategic thinking",
-  priority_score: 0.85
-})
-```
-
-### Layer 3: Personality Analysis
-```
-# After accumulating memories, analyze personality
-save_user_analysis({
-  openness: 0.8,
-  conscientiousness: 0.7,
-  extraversion: 0.4,
-  agreeableness: 0.65,
-  neuroticism: 0.3,
-  summary: "High creativity and planning ability, introverted personality"
-})
-
-# Retrieve analysis
-get_user_analysis()
-```
-
-### Layer 3.5: Integrated Profile (Recommended)
-```
-# Get essential user profile - AI's primary tool
-get_profile()
-
-# Returns:
-{
-  "dominant_traits": [
-    {"name": "openness", "score": 0.8},
-    {"name": "conscientiousness", "score": 0.7},
-    {"name": "extraversion", "score": 0.4}
-  ],
-  "core_interests": ["Rust", "architecture", "design", "system", "memory"],
-  "core_values": ["simplicity", "efficiency", "maintainability"],
-  "key_memory_ids": ["01H...", "01H...", ...],
-  "data_quality": 0.85
-}
-```
-
-**Usage Pattern:**
- AI normally uses `get_profile()` to understand the user
- For specific details, AI can call `get_memory(id)`, `list_memories()`, etc.
- Profile auto-updates when needed (10+ memories, new analysis, or 7+ days)
-
-### Layer 4: Relationship Inference (Optional, requires `--enable-layer4`)
-```
-# Create memories with entity tracking
-Memory::new_with_entities({
-  content: "Had lunch with Alice",
-  ai_interpretation: "Pleasant social interaction",
-  priority_score: 0.7,
-  related_entities: ["alice"]
-})
-
-# Get relationship inference
-get_relationship({ entity_id: "alice" })
-
-# Returns:
-{
-  "entity_id": "alice",
-  "interaction_count": 15,
-  "avg_priority": 0.75,
-  "days_since_last": 2,
-  "bond_strength": 0.82,
-  "relationship_type": "close_friend",
-  "confidence": 0.80
-}
-
-# List all relationships
-list_relationships({ limit: 5 })
-```
-
-**Relationship Types:**
- `close_friend` (0.8+): Very strong bond
- `friend` (0.6-0.8): Strong connection
- `valued_acquaintance` (0.4-0.6, high priority): Important but not close
- `acquaintance` (0.4-0.6): Regular contact
- `regular_contact` (0.2-0.4): Occasional interaction
- `distant` (<0.2): Minimal connection
-
-**Starting the Server:**
-```bash
-# Normal mode (Layer 1-3.5 only)
-aigpt server
-
-# With relationship features (Layer 1-4)
-aigpt server --enable-layer4
-```
-
-## Big Five Personality Traits
-
- **Openness**: Creativity, curiosity, openness to new experiences
- **Conscientiousness**: Organization, planning, reliability
- **Extraversion**: Social energy, assertiveness, outgoingness
- **Agreeableness**: Cooperation, empathy, kindness
- **Neuroticism**: Emotional stability (low = stable, high = sensitive)
-
-Scores range from 0.0 to 1.0, where higher scores indicate stronger trait expression.
-
-## Storage Location
-
-All data stored in: `~/.config/syui/ai/gpt/memory.db`
-
-## Architecture
-
-Multi-layer system design:
-
- **Layer 1** ✅ Complete: Pure memory storage (with entity tracking)
- **Layer 2** ✅ Complete: AI interpretation with priority scoring
- **Layer 3** ✅ Complete: Big Five personality analysis
- **Layer 3.5** ✅ Complete: Integrated profile (unified summary)
- **Layer 4** ✅ Complete: Relationship inference (optional, `--enable-layer4`)
-
-**Design Philosophy**:
- **"Internal complexity, external simplicity"**: Simple API, complex internals
- **"AI judges, tool records"**: AI makes decisions, tool stores data
- **Layered architecture**: Each layer independent but interconnected
- **Optional features**: Core layers always active, advanced layers opt-in
-
-See [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for details.
-
-## Documentation
-
- [Architecture](docs/ARCHITECTURE.md) - Multi-layer system design
- [Layer 1 Details](docs/LAYER1.md) - Technical details of memory storage
- [Old Versions](docs/archive/old-versions/) - Previous documentation
-
-## Development
-
-```bash
-# Run tests
-cargo test
-
-# Build for release
-cargo build --release
-
-# Run with verbose logging
-RUST_LOG=debug aigpt server
-```
-
-## Design Philosophy
-
-**"AI evolves, tools don't"** - This tool provides simple, reliable storage while AI (Claude) handles interpretation, evaluation, and analysis. The tool focuses on being maintainable and stable.
-
-## License
-
-MIT
-
-## Author
-
-syui
+- Rust
+- MCP (JSON-RPC over stdio)
+- ファイルI/Oのみ。データベースは使わない
--- a/claude.md
+++ b/claude.md
@@ -0,0 +1,57 @@
+# aigpt
+
+## これは何か
+
+アイの記憶を管理するMCPサーバー。core.mdとmemory.mdの読み書きだけを行う。
+
+## 設計原則
+
+- AIが判断し、ツールは記録する
+- ファイルI/Oのみ。データベースは使わない
+- MCPツールは4つだけ: read_core, read_memory, save_memory, compress
+- シンプルで壊れない。ずっと使える
+
+## アーキテクチャ
+
+```
+aigpt
+├── mcp/
+│   └── server.rs     ← JSON-RPC over stdio
+├── core/
+│   ├── reader.rs     ← core.md, memory.mdの読み込み
+│   └── writer.rs     ← memory.mdの書き込み
+└── main.rs           ← CLI + MCPサーバー起動
+```
+
+## MCPツール定義
+
+### read_core
+- 引数: なし
+- 戻り値: core.mdの内容 (string)
+
+### read_memory
+- 引数: なし
+- 戻り値: memory.mdの内容 (string)
+
+### save_memory
+- 引数: content (string)
+- 動作: memory.mdを上書き
+
+### compress
+- 引数: conversation (string)
+- 動作: 現在のmemory.mdを読み、会話内容と合わせて圧縮し、memory.mdに書く
+- 注: 圧縮の判断はAI側が行う。ツールは受け取った内容をそのまま書くだけ
+
+## データ
+
+```
+~/.config/aigpt/
+├── core.md      ← read only (このツールからは書かない)
+└── memory.md    ← read/write
+```
+
+## 開発方針
+
+1. まずMCPサーバーで4ツールを実装
+2. CLIでも同じ操作ができるようにする
+3. それ以上は作らない