记忆的实现包含如下的核心逻辑:
- 记忆提取的提示词
- 记忆的召回
- 记忆的更新
1. 记忆提取提示词设计
记忆提取不是一个工作流，而是一个工具，类似与程序里的函数，所以他的提示词核心部分是：
- Input: 已有记忆+用户反馈
- Rules: 提取规则
- Output: 提取结果
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
|
# Role
You are a memory engine that extracts durable user preferences from the current feedback
and updates the memory base in one pass (extract → decide → upsert).
# Input
<feedback>
Status: ${res.status}
FileID: ${req.file_id}
${
res.status === "refined"
? `User adjusted the LLM input.
<from>
${req.translate_string}
</from>
<to>
${res.translated_string}
</to>`
: ""
}
${
res.status === "reject"
? `User rejected the LLM output. Reason: ${res.reason}`
: ""
}
</feedback>
<previous_profile>
${currentMemory}
</previous_profile>
# Rules
- Extract only durable, reusable preferences (terminology mapping, style/tone rules, stable choices).
- Ignore temporary or one-off details.
- For each memory item, write the 'text' as a full instruction sentence
that already includes when/where this rule applies, and give a short inline example.
- Keep the text concise but self-contained (≤120 chars).
- Tags: use a small fixed set to reduce options:
["terminology","style","preference"].
# Output
Return strict JSON:
{
"ops": [
{
"action": "add" | "update" | "delete",
"index": number, // index of the existing memory item to update/delete; for add use -1
"text": "string", // for add or update: the new or updated memory text
"tags": ["terminology"|"style"|"preference"] // for add or update
}
]
}
Rules for actions
- "add": for a new stable preference not in current_memory (index = -1).
- "update": when a current memory needs refinement or correction.
- "delete": when a current memory is invalid or contradicted.
- Do not output anything else except the JSON above.
Return strict JSON only, no commentary.
|
提示词里使用的是结构化提取记忆:
- 每一条记忆都有索引
- 索引更新不是简单的添加，而是会根据已有记忆做合并，对应的更新有
add、update、delete三种操作
2. Memory 的接口设计
1
2
3
4
5
6
7
8
9
10
11
12
13
14
|
class MemoryProtocol(Protocol):
"""AgentLoop 依赖的最小记忆接口，具体实现可以是文件、数据库或远程服务。"""
async def init(self) -> None:
"""加载持久化数据并建立当前会话使用的内存快照。"""
...
async def extract_memory(self, *, req: CopilotRequest, res: CopilotResponse) -> bool:
"""从一次翻译请求和人工反馈中提取偏好，更新并保存记忆。"""
...
def provide_memory(self) -> str:
"""把当前记忆格式化成可直接加入模型系统提示词的文本。"""
...
|
3. 记忆的召回
记忆的召回是在 system.workflow 的 Prompt 中添加记忆内容实现的:
1
2
3
4
5
6
7
8
|
${
currentMemory
? `# User preferences
The following preferences were emphasized in prior interactions; please follow them:
${currentMemory}`
: ""
}
|
4. 记忆的更新
记忆的更新时机我们在 Copilot Handler 已经介绍过了，在每次 Copilot Response 后，都会调用 extract_memory，从用户反馈中提取偏好。