Save the traces! 🐳

I recently started to back up all my agent and AI traces, and I think you should too.

Why?

I have multiple personal and work accounts across many AI services and tools. ChatGPT knows about my NAS and home office setup. Claude helps me on various projects in 5 different computers. I have explored half-done side projects and research ideas with Cursor, and some implementations with Codex. I have asked Pi, Hugging Chat and ChatGPT about running 🏃‍♂️.

Two years ago these were one-off, disposable questions. I realize they are now a growing corpus of scattered and fragmented knowledge, about topics I care about. I have already found myself in a situation where I remember having bounced ideas about off Cursor, to realize this happened on an iMac I already decommissioned 🤦‍♂️.

No matter how advanced your "memory" solutions are, conversation traces are the new "file" abstraction. I want to preserve them all. I'm not fully sure how I will end up using them (some ideas later), but I do know I don't want to lose them any more.

How?

I started with something really simple:

$ for f in .claude .codex .cursor .pi; do uvx hf sync ~/${f} hf://buckets/pcuenq/traces/$(hostname -s)/${f}; done

This syncs traces and additional metadata to a private Hugging Face bucket. Ask your agent to set it up as a crontab for you, and to make sure it includes the services you use. I haven't exported my online conversations yet, that will be the next step 💪.

If you want fancier stuff, you can explore other ideas:

• Export to a dataset instead. This allows you to run queries against the data, and to see cool visualizations like this one (not one of mine, this is @mishig's 🐐). I'm pretty sure browsing, replaying and other visualizations will come to buckets at some point too.
• Use hf mount so that agents write to the bucket directly. Periodic sync is enough for me for now.

What for?

• Avoid vendor lock-in.
• Continue a project on a different computer.
• Ask your agent to summarize or consolidate various conversations.
• Extract conclusions from repeated tasks and create a Skill.
• Donate your traces so the community can train better open models.
• Reminisce with your grandchildren and their robot friends about how you steered LLMs to your liking in the good old days 👵👴🤖