Shelley's new compaction algorithm: a tool that the model calls to "squish" and "trim" prior messages.

nice viz

Why did you use to animate ?

Love your illustrarion. How can u do that?

Letting the model pick what to squish is a fun trust exercise. Does it go after giant tool outputs first? My bet is one 400-line log dump nobody needed is always the real context hog.

Keeping the pre-squish text on disk pays off: when a trim drops something needed you re-expand that one message instead of rerunning.

your second version, a summary plus the last few messages kept verbatim, is pretty much what claude code does now. what made you drop it?

like that the model picks what to trim instead of a fixed cutoff. old tool output is usually the dead weight, not the actual conversation

Failed constraints still get trimmed first

Very cool. I wonder what impact this would have on autoresearch workflows where the agent is doing heavy data analysis

the first thing a squish eats is the why: a decision gets trimmed to one line, so next session the agent reads 'no retries on webhooks' as an arbitrary constraint and proposes removing it. log every trimmed decision with its reason, outside the window. that's the whole reason i built memoryrouter.

How does it keep the prompt cache?

Letting the model pick what to drop is the interesting gamble there. Whatever decides something is safe to forget is the same thing that needs it forty turns later. Does it ever squish something and then go hunting for it again?