AI News
All projects
App

Endless-chat AI agent with a summary-tree memory, vibe-coded and open source

UniiChat is a self-hostable AI agent that keeps one chat open indefinitely and compresses its history into a tree it can zoom back into.

UniiChat by VictorTaelin · 684 GitHub stars when we found it · added

UniiChat is an AI agent that keeps one chat going indefinitely. There is no new-chat button, and the agent remembers. Its memory is a binary tree. Each message is summarized into one line, and neighboring lines are merged in pairs. Each model call sees a view of the whole chat, up to 128 KB, with recent messages in detail and older ones in broader lines. The agent can zoom back into any line, down to the original message. A cheaper model writes the summaries in the background, and the README calls it the compactor. The agent can run a shell, edit files, drive a Mac screen through the unii terminal, and spawn subagents.

The README is blunt about how the project was made. The repo is vibe-coded, with AI agents writing nearly all of it and reviewing it, and readers are told to expect bugs. It does not name the coding tools, say how long the work took, or describe how the author divided the work with the AI. The hosted version at uniichat.com sells credits. This source runs free on a self-hosted server with Bun and an Anthropic or OpenAI key. Chats are stored as JSON lines in a local folder, which the README says to back up.

In the author's words

This repo is vibe-coded. AI agents wrote nearly all of it, and AI agents reviewed it.

From the README or the author's post. We checked that the sentence is there, not that it is true.

What to take from it

If you build an agent that must remember long sessions, the pattern here is to keep cheap summaries in the prompt and the originals on disk, letting the agent expand a summary only when it needs the detail. The README credits the design of changing the view only at the end of background summarization with letting most input come from the vendor's prompt cache.

See it

AI wrote this entry from the project's own README or its author's post. That AI built the project is the author's statement, quoted here. Nobody at IKANGAI ran or tested it. How this list is made