reddit · SideProject · 09-27 22:50 50 分 ← 返回列表

Coding agents: do yours actually look up past decisions, or only when you remind them?

I used to keep a file called DECISIONS.md in every repo. Rejected approaches, why I picked the other thing, that sort of thing. Took about a month of maintaining it to realize the agent never opened it unless I told it to. I'd watch it hit the exact wall I had already documented (the CORS saga, twice) and just sit there thinking, I wrote this down, it's right there. I ended up pasting "read DECISIONS.md first" at the top of every prompt, which honestly defeated the whole purpose. I got the

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：编码 AI Agent 不会主动读取项目历史决策，导致重复踩坑、每轮会话都要人工重新交代上下文，每周浪费数小时

目标用户：重度使用 Claude Code / Cursor 等编码 Agent 的个人开发者与小型工程团队

证据：I got the thing to save me time and instead I was briefing it like a new hire every session. Hours a week, easy. ... I built something to fix this for myself. Not selling anything here, I'm just trying to figure out if it's a real problem for anyone else

变现：订阅制 SaaS：按席位或按 token 用量收费的云端团队共享记忆服务，提供免费个人版+团队付费版；或先做开源 MCP server 引流转云托管付费

MVP：2 周内做一个 MCP server：自动扫描仓库中的 DECISIONS.md、git log、CLAUDE.md 并向量化存储，通过 hook 在每次 agent 会话启动时自动检索并注入相关历史决策，无需用户手动提醒，先只支持 Claude Code 单一客户端

竞品：mem0、Zep、Cursor Memories、CLAUDE.md / AGENTS.md、各类开源 MCP memory server

摘要：发帖人在每个 repo 维护 DECISIONS.md 但发现编码 Agent 从不主动读取，导致重复踩坑并每周浪费数小时手动 briefing，于是自己做了个基于 MCP 的云端记忆层，统一保存决策、失败尝试、会话和技能供所有 agent 共享。他明确表示不卖东西，只是在验证这是否是普遍痛点，并追问他人用了什么 hooks 或系统提示来让 agent 可靠回忆。痛点真实且具体，但尚无任何付费意愿或收入数据，且 memory 赛道已有 mem0、Zep、Cursor 内置记忆等众多竞品。