Xia Chen

GNI Postdoctoral Fellow @TUM | Human-AI Alignment | Building ThoughtDAG, an open-source canvas where wires are the context.

[Xia Chen @ChatchanX]

AI chat hides its most important state: what the model is actually seeing.

ThoughtDAG makes that context visible and editable. Every turn is a node. Wires decide what reaches the next answer.

Delete a wire. Change the answer.

https://t.co/IisYjzikzF

[Xia Chen @ChatchanX]

@priestessofdada Signed immutable collaboration facts feel right. For deliberate agent memory, how do you stop a validly signed but stale fact entering a future run? Is “superseded” another signed fact, with consumers resolving policy at read time rather than rewriting history?

[Xia Chen @ChatchanX]

@rohanpaul_ai Useful distinction. I’d add a third boundary at inference time: durable knowledge can exist in the Hub without entering every prompt. Otherwise “shared memory” quietly becomes permission to influence every task. Does Cohub review storage and per-run retrieval separately?

[Xia Chen @ChatchanX]

Can you still see an AI agent's context after its terminal closes?

ThoughtDAG Session Atlas mirrors local Codex and Claude Code sessions as maps, turns, tool calls, and results included.

Reorganize. Condense what matters. Continue with the context you choose.

Link in reply.

[Xia Chen @ChatchanX]

@soobadina Thank you! Heartfully invite you to try it! Any feedback is welcome :)

[Xia Chen @ChatchanX]

ThoughtDAG v0.4.1 is live — our biggest update: Session Atlas.

Local Codex + Claude Code sessions shouldn’t stay buried in terminals.

Open them as live, editable context graphs. In one real session, I condensed 201,400 tokens to 15,200 — then continued with only what mattered.

[Xia Chen @ChatchanX]

This research summary kept talking about dinner.

Same prompt. Same model. I deleted one context edge and regenerated.

The dinner advice disappeared.

That edge wasn’t decoration. It was model input.

ThoughtDAG is open source:
https://t.co/IisYjzikzF

[Xia Chen @ChatchanX]

@johnzabroski Phoenix looks strong for runtime traces. My workflow is manual: hold prompt, model, and settings fixed, change one context dependency, then compare the rerun. I’m building this as an editable DAG; pairing it with reusable trace components feels like the missing bridge.

[Xia Chen @ChatchanX]

@soobadina A practical starting point: keep private files and long-term notes local, send only the minimum task-specific excerpt to the cloud, and make that handoff explicit. If a tool does not show exactly what leaves your machine, assume the whole chat context may be sent.

[Xia Chen @ChatchanX]

@QingQ77 Hi Geek Lite — do you accept open-source project submissions? I’d love to share one that may fit your audience.

[Xia Chen @ChatchanX]

ThoughtDAG is starting to travel.

Now in the official Homebrew Cask, awesome-opensource-ai, awesome-knowledge-management, and Local First News.

An open-source canvas where wires control what the model sees.

https://t.co/IisYjzikzF

[Xia Chen @ChatchanX]

@ainewsusa Exactly. “Human in the loop” needs a concrete control surface: people need to see what evidence and prior decisions shaped a recommendation before they can meaningfully intervene.

[Xia Chen @ChatchanX]

@AhmadBaracat That parallel feels useful. A key design question is whether the human correction becomes a durable, inspectable part of the future decision path—or just another opaque training trace.

[Xia Chen @ChatchanX]

@aakashbuildsAI Agreed. Privacy is not only about where inference runs; it is also about being able to inspect which earlier notes, files, and decisions are still influencing the next request.

[Xia Chen @ChatchanX]

Invited a few people to use ThoughtDAG for real tasks.

Their exploration paths looked completely different: some went deep, some branched widely, some kept dead ends out of the next context.

Not a personality test. Traces of how people work with AI.

Want to contribute one?

[Xia Chen @ChatchanX]

@Nemo001ai Yes, the missing unit of work is often the change, not the final answer. If an assumption changes, we should see which later decisions inherited it, then choose what to preserve, revise, or recompute. That makes a workspace more than an agent launcher.

[Xia Chen @ChatchanX]

@DavidOndrej1 The most underrated part of a local AI stack is being able to see what is running and why. Control beats magic here.

[Xia Chen @ChatchanX]

@bozhou_ai The approval gate is not merely a safety pause; it is also where teams retain the rationale and context behind a change. Without that trace, faster generation can make review harder.

[Xia Chen @ChatchanX]

@jinglian Useful comparison. For local workflows, the decisive friction is often not raw tokens per second, but whether model choice, source context, and state stay inspectable and portable.