sync cleaned docs: dsh-plugin/memory-director/README.md
Browse files
dsh-plugin/memory-director/README.md
ADDED
|
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# dsh-memory-director
|
| 2 |
+
|
| 3 |
+
**MemoryDirector 插件** —— 给 DeepSeek Harness 加"长期记忆"能力。
|
| 4 |
+
|
| 5 |
+
官方 Harness 的 compaction 只做**摘要压缩**，没有任何"跨会话记住重要事实"的概念。
|
| 6 |
+
这个插件补上 AgentFrame 的 MemoryDirector：
|
| 7 |
+
|
| 8 |
+
- **每轮对话后**：LLM 分析本轮，决定记住什么 / 忘记什么
|
| 9 |
+
- **每次请求前**：检索相关记忆，注入模型上下文
|
| 10 |
+
- **持久化**：记忆存 JSON 文件（`~/.dsh/memory.json`）
|
| 11 |
+
|
| 12 |
+
## 为什么是差异化
|
| 13 |
+
|
| 14 |
+
| 能力 | 官方 Harness | dsh-memory-director |
|
| 15 |
+
|------|-------------|---------------------|
|
| 16 |
+
| 上下文压缩 | compaction | (可配合) |
|
| 17 |
+
| **记住重要事实** | × 没有 | **LLM 决策** |
|
| 18 |
+
| **跨会话记忆** | × | 持久化 |
|
| 19 |
+
| **记忆检索注入** | × | pre-step 注入 |
|
| 20 |
+
| **遗忘管理** | × | 重要度衰减 |
|
| 21 |
+
|
| 22 |
+
## 接入
|
| 23 |
+
|
| 24 |
+
```yaml
|
| 25 |
+
# profile 的 cordis.patch.yml
|
| 26 |
+
- id: memory-director
|
| 27 |
+
name: '@agentframe/dsh-memory-director'
|
| 28 |
+
config:
|
| 29 |
+
provider: deepseek-official
|
| 30 |
+
model: deepseek-v4-flash
|
| 31 |
+
maxTokens: 256
|
| 32 |
+
dedupThreshold: 0.8
|
| 33 |
+
forgetThreshold: 0.1
|
| 34 |
+
storePath: ~/.dsh/memory.json
|
| 35 |
+
auto: true
|
| 36 |
+
```
|
| 37 |
+
|
| 38 |
+
## 机制
|
| 39 |
+
|
| 40 |
+
```
|
| 41 |
+
agent/turn-stopping (每轮结束)
|
| 42 |
+
→ 提取本轮文本 → LLM 决策 {remember, forget, importance}
|
| 43 |
+
→ remember → 存入记忆库 (去重)
|
| 44 |
+
→ forget → 删除过时记忆
|
| 45 |
+
|
| 46 |
+
agent/pre-step (每次请求前)
|
| 47 |
+
→ 取最后用户消息 → 检索相关记忆
|
| 48 |
+
→ 注入 system 消息 (memory-director block)
|
| 49 |
+
→ 模型可基于记忆回答
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
## 配置
|
| 53 |
+
|
| 54 |
+
| 字段 | 默认 | 说明 |
|
| 55 |
+
|------|------|------|
|
| 56 |
+
| `provider` | deepseek-official | 决策用 provider |
|
| 57 |
+
| `model` | deepseek-v4-flash | 决策用模型 (便宜快速) |
|
| 58 |
+
| `maxTokens` | 256 | 决策输出上限 |
|
| 59 |
+
| `dedupThreshold` | 0.8 | 记忆去重阈值 |
|
| 60 |
+
| `forgetThreshold` | 0.1 | 遗忘阈值 (重要度衰减) |
|
| 61 |
+
| `storePath` | ~/.dsh/memory.json | 持久化路径 |
|
| 62 |
+
| `auto` | true | 自动挂钩 agent 循环 |
|
| 63 |
+
|
| 64 |
+
## License
|
| 65 |
+
|
| 66 |
+
GPL-3.0 © Cloud LTE Studio / AgentFrame
|