Gnosh.ai for "Functional AI"; functional context management, no compaction

It replays a real Gnosh run: Qwen3.8 27B on a desktop AMD box (Ryzen AI Max+ 395) renaming a small Go project from klash to gnosh over two sessions, 31 October 2026. All six of the task's checks passed.

Every call, label, result and thought shown is from that run. Each block's height is that call's real size, at about 3.5 characters per token.

The window is shrunk to 2,400 tokens of room so a small task fills it, and the summary blocks are drawn scaled to it. The real run had 262K and needed neither a summary nor a slide.

The ordinary lane replays the same calls with a plain todo list instead of mem, and stays in one session, which is the easier path for it. Its summaries are illustrative.

The instructions and tools (about 15K tokens) sit under both columns, not to scale. They stay cached either way.

Measured, from the Gnosh README: in a real run's server log, the request after a pop processed 2,572 tokens instead of the whole 20K-token conversation.

In Head to head the two lanes race on GPU time, played in about 12 seconds of real time whatever the GPU time comes to (the clock runs at roughly 80×): each side's timer is how long the model has been working on the box's GPU, both run on the same clock, and each stops the moment its side finishes the task. Each step costs its tokens read at 179 tokens/s and written at 13.3 tokens/s: the box's median speeds over 4,122 requests in its llama-server log (Qwen3.8 27B, the 36 hours to 31 October 2026, 20:00 UTC). A summary costs the 1,244-token summary instructions read, 4,504 tokens written and re-read (the median of 24 real summaries in this box's own session databases; the largest was 13,786), so about six minutes each on this box; a slide costs the tokens after the cut re-read; a new session's first prompt carries a mem block of about 250 tokens. Tool time is left out, since the same tools run in both lanes; mem's own time is the run's receipts ("ok · mem · ran 47ms"). The ordinary agent's todo-list updates are sized like a typical todo tool call, about 120–170 tokens written each. "Task completed" counts the five parts of the rename (code, Makefile, config, README, tests) at the step where each was really finished in the run; the README todo itself stayed held in mem, as chapter 2 shows.

In the real run the second session renamed the README while verifying, so the README todo itself stayed open.

An AI agent works by sending the model everything it has seen so far, on every step. Here is what that looks like today, and what Gnosh does instead, on the same real task.

space pauses · ← → change chapter · click a chapter to jump