Watch a context window breathe

This is a real client-side simulation of our compaction engine. Pick a model size, feed the window messages, then hit Compact and watch the policy reclaim space — pinned messages protected, tool outputs summarized, low-priority content evicted oldest-first.

1 · Choose a model

2 · Add messages

3 · Watch it fill, then compact

Policy: summarize tool outputs → evict low-priority oldest-first until usage < 60%. Pinned and high-priority messages are never evicted.

Message log (newest first)

What the demo simulates

The same three-phase policy the production engine runs: summarize verbose tool outputs in place, then evict by priority tier (oldest first), stopping at a configurable target utilization. Pinned content is provably untouched.

What it doesn't simulate

Semantic recall (retrieving evicted content on demand), cross-session memory, and multi-agent pools need a backend — that's what the platform adds. This page is pure client-side JavaScript.

Try this

Load the sample, switch to the 8K model, and watch overflow. Then compact: tool outputs collapse to ~22% of their size and the window drops back under 60% — with the system prompt and user intent intact.

Like what you saw?

The demo runs on toy data. The platform runs on your agents — with recall, observability, and policies you can tune.