Hierarchical Memory System

TAISUN's hierarchical memory architecture based on Mem0 research, providing 26% accuracy improvement through structured memory layers.

Architecture Overview

┌─────────────────────────────────────────────────────────────────┐
│                    HIERARCHICAL MEMORY                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐ │
│  │   SHORT-TERM    │  │    LONG-TERM    │  │    EPISODIC     │ │
│  │   (Session)     │  │   (Persistent)  │  │   (Events)      │ │
│  ├─────────────────┤  ├─────────────────┤  ├─────────────────┤ │
│  │ taisun-proxy    │  │ Qdrant Vector   │  │ claude-mem      │ │
│  │ InMemoryStore   │  │ Database        │  │ Observations    │ │
│  ├─────────────────┤  ├─────────────────┤  ├─────────────────┤ │
│  │ TTL: Session    │  │ TTL: Permanent  │  │ TTL: 30 days    │ │
│  │ Size: 100 items │  │ Size: Unlimited │  │ Size: 50/day    │ │
│  │ Search: Token   │  │ Search: Vector  │  │ Search: ID/Time │ │
│  └─────────────────┘  └─────────────────┘  └─────────────────┘ │
│           │                   │                    │            │
│           └───────────────────┼────────────────────┘            │
│                               ▼                                 │
│                    ┌─────────────────┐                         │
│                    │  MEMORY ROUTER  │                         │
│                    │  (Consolidation)│                         │
│                    └─────────────────┘                         │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Memory Layers

1. Short-Term Memory (Working Memory)

System: taisun-proxy InMemoryStore Purpose: Current session context

| Property | Value | |----------|-------| | Storage | In-memory | | TTL | Session duration | | Max Items | 100 | | Search | Token-based | | Use Cases | Current task context, recent commands, temp data |

# Store in short-term
memory_add type="short-term" content="現在のタスク: API実装"

# Retrieve
memory_search query="タスク"

2. Long-Term Memory (Semantic Memory)

System: Qdrant Vector Database Purpose: Persistent knowledge and patterns

| Property | Value | |----------|-------| | Storage | Qdrant (localhost:6333) | | TTL | Permanent | | Max Items | Unlimited | | Search | Vector similarity | | Use Cases | Code patterns, learned solutions, domain knowledge |

# Store important pattern
qdrant-store text="認証にはJWTを使用し..." metadata={topic: "auth"}

# Semantic search
qdrant-find query="認証の実装方法"

3. Episodic Memory (Event Memory)

System: claude-mem Observations Purpose: Decision history and context trails

| Property | Value | |----------|-------| | Storage | JSONL files | | TTL | 30 days | | Max Items | ~50/day | | Search | ID, timestamp, type | | Use Cases | Past decisions, debugging context, learning history |

# Auto-captured by hooks
# Access via MCP
mcp__claude-mem-search__search query="bugfix"
mcp__claude-mem-search__timeline date="2026-01-19"

Memory Flow

Information Lifecycle

1. CAPTURE (Short-Term)
   User input → Session context → Working memory

2. CONSOLIDATE (Short → Long)
   Important patterns → Vector embedding → Qdrant storage

3. OBSERVE (Episodic)
   Decisions, discoveries → claude-mem → Timestamped records

4. RETRIEVE (All Layers)
   Query → Router → Best matching layer → Response

Consolidation Rules

| Trigger | Action | |---------|--------| | Session end | Important short-term → Long-term | | Pattern detected | Auto-store in Qdrant | | Decision made | Log to episodic | | Error resolved | Store solution in long-term |

Usage Patterns

1. Remember Important Information

User: このAPIパターンを覚えておいて
      [code snippet]

AI: 1. Short-term に即座に保存
    2. 重要度判定（コードパターン = HIGH）
    3. Qdrant に永続化
    4. claude-mem に観察記録

2. Retrieve Past Knowledge

User: 以前話した認証の実装方法は？

AI: 1. Qdrant でセマンティック検索
    2. claude-mem でエピソード検索
    3. 関連情報を統合
    4. コンテキスト付きで回答

3. Learn From Session

# Session end hook automatically:
1. Extracts key decisions
2. Stores successful patterns
3. Records errors and solutions
4. Updates long-term memory

Performance Benefits (Mem0 Research)

| Metric | Improvement | |--------|-------------| | Accuracy | +26% | | P95 Latency | -91% | | Token Usage | -90% |

Source: Mem0 Research Paper

Integration Points

With Existing TAISUN Systems

| System | Integration | |--------|-------------| | taisun-proxy | memory_add, memory_search tools | | Qdrant MCP | qdrant-store, qdrant-find tools | | claude-mem | Auto-observation hooks | | SessionStart | State injection | | SessionEnd | Memory consolidation |

With Other MCPs

# Context7 + Long-Term Memory
「use context7 でReact 19の新機能を学習して、覚えておいて」

# GPT Researcher + Memory
「市場調査して、重要なポイントを長期記憶に保存」

Best Practices

Explicit Memory Commands

✅ 「これを長期記憶に保存して」
✅ 「前回のセッションで話した〇〇について」
❌ 「覚えておいて」（曖昧）

Tag Important Information

metadata: { topic: "auth", type: "pattern", priority: "high" }

Regular Memory Cleanup

Outdated patterns should be removed from long-term memory

Trust the Consolidation

Let auto-hooks handle session → long-term migration

hierarchical-memory