I have been building software with coding agents. They can write a feature, debug a nasty failure, and then, in a fresh chat, forget the reason we rejected an approach yesterday. Suddenly I am the human database. That is bullshit.
A handoff document helps, but it cannot predict which tiny correction will matter three weeks from now. Stuff every conversation into a prompt and the new chat becomes slow and expensive. Ask the agent to search its own history and the answer depends on what that particular app can access.
So I built Lore, a local MCP memory tool for coding projects. The current alpha is free and Windows-only.
What it actually does
Lore reads accessible local Codex, Claude Code, and Qwen Code conversations into an archive on your PC. In a fresh Codex task or Claude Desktop chat, you can ask a question like: “Use Lore. Why did we reject the first deployment plan?” It retrieves relevant excerpts with conversation and message IDs. If the answer matters, you can inspect the cited turn and its surrounding context.
That citation is the point. A smooth answer without evidence is just a confident story. If two old chats disagree, I want to see the disagreement instead of getting a made-up synthesis.
A test that is harder to bullshit
Pick a decision from an old local coding-agent chat. Choose one whose answer you know, but which you never wrote into a project document. Open a fresh Codex task or Claude Desktop chat. Ask the question without giving away the answer, then check the cited source. Did Lore find the right turn? Does the nearby context actually support what the agent said?
If it misses, I want to hear about that too. A believable wrong answer is worse than an honest “I couldn't find it.”
The boundaries, because they matter
Lore does not magically see every chat on the internet. Installing the MCP does not automatically import ordinary ChatGPT or Claude cloud conversations. It reads supported local coding-agent histories that are accessible on your machine. It does not sync between PCs. It is a closed-source alpha, and the Windows binary is unsigned, so Windows may warn about it. The archive stays local, but any excerpts you send to an AI client are subject to that client's policies.
The setup guide is here. If you try it, tell me three things: did setup work, did the first useful question get answered, and was the citation right?
What does your current handoff system keep losing?
Top comments (0)