Temporarily addressing the issue of high chat costs by limiting the number of historical messages the AI remembers / 通过限制AI记忆的历史消息条数来暂时解决聊天费用过高的问题

[AdventuringTime]

I've been chatting for the past two days, and noticed that the number of tokens consumed for input keeps increasing, leading to higher costs. As a student, it's hard for me to afford such expenses. So, I dug up the source code and temporarily limited the number of historical messages included in each AI response to control costs.

我最近聊了两天，发现输入消耗的 tokens 越来越多，产生的费用也越来越多，但是学生党难以支付那么多费用。于是我把源代码翻了出来，暂时限制每次 AI 回答时所输入的历史消息条数，来控制费用消耗。

This is just a temporary workaround before implementing a memory system. I haven't added a feature that allows users to set the maximum number of historical messages, so I won't be submitting a PR. Instead, I'm sharing this in the Discussion for reference, hoping it might help others facing similar issues~

不过这只是作为记忆系统实现前的权宜之计，而且我也没写让用户设置最大历史消息条数的部分，所以就不提交 PR 啦，只在 Discussion 分享一下，以供遇到类似问题的同好参考~

Solution: Add the following code in packages\stage-ui\src\stores\chat.ts and modify the subsequent variable:

解决方案：在packages\stage-ui\src\stores\chat.ts中添加以下代码，并修改随后的一处变量：

          },
        ],
      })
// Original codes above / 以上为原始代码

      // 限制历史消息条数为 60 条（保留系统消息 + 最近的非系统消息）
      const MAX_HISTORY_MESSAGES = 60

      // 从会话中获取当前消息列表的副本
      let messagesToSend = sessionMessagesForSend.slice()

      if (messagesToSend.length > MAX_HISTORY_MESSAGES) {
        // 分离系统消息和其他消息
        const systemMessage = messagesToSend.find(msg => msg.role === 'system')
        const otherMessages = messagesToSend.filter(msg => msg.role !== 'system')

        // 系统消息存在时，保留系统消息 + 最近的 (MAX_HISTORY_MESSAGES - 1) 条非系统消息
        if (systemMessage) {
          const recentOthers = otherMessages.slice(-(MAX_HISTORY_MESSAGES - 1))
          messagesToSend = [systemMessage, ...recentOthers]
        } else {
          // 如果没有系统消息，则直接保留最近的 MAX_HISTORY_MESSAGES 条消息
          messagesToSend = otherMessages.slice(-MAX_HISTORY_MESSAGES)
        }
      }
// Original codes below / 以下为原始代码

      let newMessages = MessagesForSend.map((msg) => {
        // 此处将 sessionMessagesForSend 改为 messagesToSend
        const { context: _context, ...withoutContext } = msg
        const rawMessage = toRaw(withoutContext)

        if (rawMessage.role === 'assistant') {
          const { slices: _slices, tool_results, categorization: _categorization, ...rest } = rawMessage as ChatAssistantMessage

I don't know TypeScript programming, and the above code was modified with the help of AI. Please forgive any mistakes.

本人并不了解 Typescript 编程，以上代码由AI辅助修改，请各位大佬见谅。

This English translation is generated by AI, please forgive any grammatical issues.

英文由AI翻译生成，如有语法问题请见谅。

[AdventuringTime]

I think, compared to long-term and short-term memory, this cost is still high. Looking forward to the implementation of long-term and short-term memory soon.

感觉相比长短期记忆来说，这成本还是有些高了。期待长短期记忆即将实装。

[0xSelenicDove]

我之前写过一个关于短期记忆的 pr，不过其实挺简单粗暴，就是只能回想最近几条消息： #1707
然后之前也有不少关于长短期记忆的想法，有兴趣可以去 PR 那里看看！