Here's a CLI tool that lets agents search over Claude Code or Codex session transcripts. This query would be very difficult without it:

"have we worked on X before?"

Typically, agents would need to grep over local repos or prev JSONL transcripts, or hope that it was written to a .md file indexed at the start of its context.

`traces search <query>`

<query> is embedded and similar user prompts embeddings or tool_call embeddings under the same prompt is surfaced across all session transcripts.

`traces search <query>` itself becomes a tool call, so in theory, future searches for <query> become cached.

As much as I love Karpathy, LLM Wiki doesn't seem to work for agents. I think they prefer to use tools.