Re:Chroma Context-1: Training a Self-Editing Search Agent
by nostrebored on Thursday March 26, 2026 @11:48PM
(#47538939)
Is there a reason to prune individually instead of introducing a tombstoning approach like kimi?
Most of my harnesses around agentic retrieval wind up implementing the poor man’s version of this via isolated context windows and recursion. But it seems like an entire trajectory is more likely to be erroneous than its docs, and you could just rewrite true positives in poor trajectories as the summary?
Re:Chroma Context-1: Training a Self-Editing Search Agent
by d0963319287 on Friday March 27, 2026 @02:09AM
(#47539545)
There’s an interesting pattern here.
When the context gets edited and compressed enough,
it sometimes stops behaving like something that needs to be managed,
and starts reconstructing what it needs automatically.
At that point, the system feels less like search + filtering,
and more like a single process that’s continuously re-forming its own context.
Re:Chroma Context-1: Training a Self-Editing Search Agent
by Xmd5a on Friday March 27, 2026 @09:34AM
(#47542473)
> When the context gets edited and compressed enough, it sometimes stops behaving like something that needs to be managed, and starts reconstructing what it needs automatically.
Re:Chroma Context-1: Training a Self-Editing Search Agent
by lotteseifert on Friday March 27, 2026 @02:30AM
(#47539633)
We published our research in December and told Chroma's CEO Jeff. 4 months later, Chroma republished it without citing it. We think this sets a pretty bad precedent:
Re:Chroma Context-1: Training a Self-Editing Search Agent
by redwood on Friday March 27, 2026 @09:24AM
(#47542362)
Out of curiosity did you reach out privately to explain you were surprised by their omission, and give them a change to update things? or did you go straight into the public story about copycats not giving credit where credit is due? Because it seems very conceivable to me that in the game of telephone of getting something out there some things may have been lost that might have been added in if you had reached out in good will. If you had, and they still ignored, that would be a different story entirely.
Re:Chroma Context-1: Training a Self-Editing Search Agent
by jghiglia on Tuesday March 31, 2026 @08:04PM
(#47595106)
The self-editing angle is interesting but the part that breaks in production is what happens when the search agent's memory of previous searches is wrong or stale. We ran into this building multi-agent systems — the agent confidently re-runs searches it already ran because it has no reliable memory of what it tried. The self-editing loop helps but only if the memory layer underneath it is trustworthy.
Most of my harnesses around agentic retrieval wind up implementing the poor man’s version of this via isolated context windows and recursion. But it seems like an entire trajectory is more likely to be erroneous than its docs, and you could just rewrite true positives in poor trajectories as the summary?
When the context gets edited and compressed enough, it sometimes stops behaving like something that needs to be managed, and starts reconstructing what it needs automatically.
At that point, the system feels less like search + filtering, and more like a single process that’s continuously re-forming its own context.
Do you have examples?
https://x.com/maxrumpf/status/2037365748973384154