【RFC】OpenViking接入openclaw记忆系统

[Mijamind719]

openviking接入openclaw记忆系统方案讨论：
openclaw关联PR openclaw/openclaw#14073
openViking关联pr：
记忆入库的语言期望与用户语言保持一致，提升召回历史的相关性：#137

[MaojiaSheng]

还有一个项目叫 memoh 可以参考下

[Mijamind719]

OpenViking接入openclaw记忆系统设计结构图：

会话前注入相关历史信息：

会话结束总结入库，openviking存储信息查询：

[evpeople]

我在 Astrbot （一个类似openclaw的项目，可能提供了更丰富的hook） 中，也在尝试使用openviking做长期记忆。下面的是我的方案，希望有可以参考的地方

存储

在对话开始的时候，同时启动一个session，然后将用户发的所有内容，和LLM实际发给用户的信息，都在对应的hook中存到openviking里面。

查询

在接收到用户信息的hook中

首先对用户信息进行一次关键词匹配，如果匹配上了如“之前”，“还记得”，“上次”等表示回忆的内容时 使用 search 进行一次精准搜索。

其他情况下，自动以用户消息来进行一次 find 查询，并返回对应的内容的 uri 供 llm 后续需要更明确的内容时，使用 uri 作为参数查询。

不默认使用search 主要是考虑search 本身也会使用 llm ，造成整个对话的延时的大幅增加。

记忆提交

session的commit我的设计中有多个时机

1. x轮对话后自动commit
2. x分钟内会话没有新消息后，自动commit
3. 在上下文压缩的hook中，自动commit

提供给 agent 的工具

仅提供读取工具给 agent ，主要是 ls read tree search，没有考虑好是否应该提供abstract等工具给 agent

对于垃圾记忆

如果搜的足够准确，那么垃圾记忆除了占点硬盘空间外，没有其他任何影响，所以我下掉了最开始开发的对垃圾记忆清理的agent

In Astrbot (a project similar to openclaw, possibly offering more extensive hooks), I am also experimenting with using OpenViking for long-term memory. Below is my approach, hoping it may offer some useful references.

Storage

At the beginning of a conversation, a session is started simultaneously. All content sent by the user and the actual information sent by the LLM to the user are stored in OpenViking via the corresponding hooks.

Query

In the hook that receives user messages:

First, perform a keyword match on the user's input. If matches are found for terms like "before," "remember," or "last time" (indicating recall-related content), a precise search is conducted using the search function.

In other cases, a find query is automatically performed based on the user's message, and the URI of the corresponding content is returned. This allows the LLM to use the URI as a parameter for more specific queries later if needed.

The reason search is not used by default is that it involves the LLM, which could significantly increase the latency of the entire conversation.

Memory Submission

The session commit is triggered at multiple points in my design:

1. Automatically after X rounds of conversation.
2. Automatically after no new messages in the session for X minutes.
3. Automatically in the hook for context compression.

Tools Provided to the Agent

Only read tools are provided to the agent, mainly ls, read, tree, and search. It is still under consideration whether tools like abstract should be provided to the agent.

Handling Junk Memory

If the search is accurate enough, junk memory has no impact other than taking up some hard drive space. Therefore, I have removed the agent initially developed for cleaning up junk memory.

[Mijamind719]

a related PR #307 has been merged
提供以下方式完成安装部署：
1.可以使用agent完成安装：
使用方式例如对openclaw说：查看OpenViking/examples/openclaw-memory-plugin/INSTALL-ZH.md at main · volcengine/OpenViking 这篇安装部署指导文档，帮我完成openclaw+openviking的插件安装以及启动

2.用户可以自主参考安装部署指导文档完成安装
OpenViking/examples/openclaw-memory-plugin/INSTALL-ZH.md at main · volcengine/OpenViking
环境准备
组件 版本要求
Python >= 3.10
Node.js >= 22
cmake —
g++(gcc-c++) —
Go >= 1.25 编译 AGFS 服务端（仅 Linux 源码安装）
openclaw
环境准备完成之后可以执行以下命令完成openclaw+openviking插件的安装
git clone GitHub - volcengine/OpenViking: OpenViking is an open-source context database designed specifically
cd OpenViking
npx ./examples/openclaw-memory-plugin/setup-helper

[MaojiaSheng]

Note: If OpenViking Server is already deployed on cloud, please say: "Now we need to use OpenViking plugin and CLI, please clone OpenViking (https://github.com/volcengine/OpenViking), and then 'npx ./examples/openclaw-memory-plugin/setup-helper' to install your plug-in"