Memory Sync Enhanced

增强版记忆系统 - Ebbinghaus 遗忘曲线 + Hebbian 共现图

Security Scan

The skill's code, instructions, and resource accesses are coherent with a local memory / co‑occurrence tracker: it creates a local SQLite DB under the user's config dir and contains no network calls or unexpected credential requests.

Purpose & Capability

Name and description describe a memory system combining decay + co-occurrence. The included Python tracker implements Hebbian co-occurrence logic and a local SQLite DB at ~/.config/cortexgraph/co_occurrence.db — this is consistent with the stated purpose and no unrelated capabilities or credentials are requested.

Instruction Scope

SKILL.md instructs the agent to run various shell scripts (sync-memory.sh, search.sh, gc.sh, etc.). Those scripts are not included in the package; the provided Python file is limited to recording/querying co-occurrence and printing stats. The instructions do not ask the agent to read unrelated system files or exfiltrate data, but you should be aware the skill expects other scripts to exist (not provided here).

Install Mechanism

No install spec; the skill is instruction-only plus one harmless Python script. Nothing is downloaded from external URLs and no archives are extracted. Risk from install mechanism is low.

Credentials

No environment variables, credentials, or external endpoints are required. The script only writes to a user-scoped path (~/.config/cortexgraph), which is proportional to a local memory store.

Persistence & Privilege

always is false and the skill does not request system-wide changes. The only persistent artifact is a local SQLite DB in the user's config directory. The skill does not modify other skills' configs or request elevated privileges.

Assessment

This skill appears internally consistent and implements a local co-occurrence tracker. Before installing or running: (1) inspect any shell scripts referenced in SKILL.md (sync-*.sh, search.sh, gc.sh) because they are not included here — don't run unknown shell scripts; (2) note the tracker will create a database at ~/.config/cortexgraph/co_occurrence.db and directories under ~/.config — back up or review that location if you have naming conflicts; (3) if you plan to store sensitive memories, consider encrypting that DB or restricting filesystem permissions; (4) if you expect network integration (CortexGraph remote services), verify those components separately — this package makes no network calls itself.

增强版记忆系统

结合 Ebbinghaus 遗忘曲线 + Hebbian 共现图 的双层记忆架构。

架构

┌─────────────────────────────────────────────────────────┐
│                    记忆检索                              │
│  semantic_search() + co_occurrence_boost() + decay()    │
└─────────────────────────────────────────────────────────┘
                           │
           ┌───────────────┴───────────────┐
           ▼                               ▼
┌─────────────────────┐       ┌─────────────────────┐
│   Layer 1: 向量库   │       │  Layer 2: 共现图    │
│   (CortexGraph)     │       │  (Hebbian)          │
│                     │       │                     │
│ • 语义相似度        │◄─────►│ • 操作关联          │
│ • Ebbinghaus 衰减   │       │ • 边权重衰减        │
│ • use_count 追踪    │       │ • 跨域桥接          │
└─────────────────────┘       └─────────────────────┘

核心算法

Layer 1: Ebbinghaus 遗忘曲线

score = (use_count)^β × e^(-λ × Δt) × strength

• β = 0.6（使用频率权重）
• λ = ln(2) / half_life（默认 3 天）
• strength = 1.0-2.0（重要性）

Layer 2: Hebbian 共现图

effective_weight = weight × 2^(-age_days / 30)

• 每次记忆 A 和 B 同时被检索 → 边(A,B) 权重 +1
• 边权重 30 天半衰期
• 跨域桥接：音乐记忆 ↔ 编码记忆（因为同时发生）

检索流程

def retrieve_memory(query, top_k=10):
    # 1. 语义搜索
    semantic_results = cortexgraph.search(query, top_k * 2)
    
    # 2. 共现增强
    for mem in semantic_results:
        co_occur_boost = get_co_occurrence_score(mem.id, recent_context)
        mem.boosted_score = mem.semantic_score + co_occur_boost * 0.3
    
    # 3. 遗忘曲线过滤
    for mem in semantic_results:
        mem.final_score = mem.boosted_score * mem.decay_factor
    
    # 4. 返回 Top K
    return sorted(semantic_results, key=lambda x: x.final_score)[:top_k]

记忆类型

STM (短期记忆)

• JSONL 格式
• 快速读写
• 高衰减率（3天 half-life）
• 存储日常日志

LTM (长期记忆)

• Obsidian Markdown
• 永久存储
• 低衰减率（30天 half-life）
• 存储重要洞察

Co-occurrence Graph

• SQLite 边表
• 30天 half-life
• 记录记忆之间的关联

数据结构

CortexGraph 记录

{
  "id": "uuid",
  "content": "记忆内容",
  "embedding": [0.1, 0.2, ...],
  "use_count": 5,
  "last_used": "2026-02-19",
  "strength": 1.5,
  "created_at": "2026-02-15",
  "tags": ["daily-log", "finding"]
}

Co-occurrence 边

CREATE TABLE co_occurrence (
  memory_a TEXT,
  memory_b TEXT,
  weight REAL,
  last_updated TEXT,
  PRIMARY KEY (memory_a, memory_b)
);

使用方法

同步记忆

# 同步 MEMORY.md
./scripts/sync-memory.sh

# 同步每日日志
./scripts/sync-daily.sh 2026-02-19

# 记录共现
./scripts/record-co-occurrence.sh

检索记忆

# 语义搜索
./scripts/search.sh "量化交易"

# 增强搜索（语义 + 共现）
./scripts/search-enhanced.sh "量化交易"

记忆管理

# 查看记忆统计
./scripts/stats.sh

# 垃圾回收（删除低分记忆）
./scripts/gc.sh --threshold 0.1

# 晋升到长期记忆
./scripts/promote.sh <memory_id>

统计示例

=== 记忆系统统计 ===

总记忆数: 2,400
共现边: 803 (连接 366 个记忆)
平均每个记忆连接: 2.2 个

记忆分布:
- STM: 1,800 (75%)
- LTM: 600 (25%)

衰减状态:
- Danger zone (0.15-0.35): 120 个
- Healthy (0.35-0.65): 1,500 个
- Strong (>0.65): 780 个

与其他系统对比

系统	向量搜索	遗忘曲线	共现图
Markdown 文件	❌	❌	❌
CortexGraph 原版	✅	✅	❌
Zeph 的 Hebbian	✅	❌	✅
本系统	✅	✅	✅

设计理念

1. 遗忘是功能 - 不是所有记忆都需要永久保存
2. 关联即记忆 - 两个记忆同时出现 = 它们有关联
3. 跨域桥接 - 穿衣服记录和调试记录可以关联
4. 个性在桥接中 - 跨域边是 personality 所在

参考

• CortexGraph
• @Zeph 的 Hebbian 共现图帖子 (The Colony)
• Ebbinghaus 遗忘曲线理论

---

版本: 2.0.0 结合 Ebbinghaus 遗忘曲线 + Hebbian 共现图