01 / CONVERSATION HISTORYWaiting
Baseline
The question and all seeded conversation turns.
—input tokens
Output —Cached input —
MODEL ANSWER
The live answer will appear here.
PROJECT MEMORY, ONE COMMIT AT A TIME
Remember the decision.
Carry less history.
Give Qwen the same question twice: once with a seeded conversation, then with a relevant StateTree commit note. Compare the input tokens the model actually reports.
THE FIXED EXAMPLE
What is the agreed export retention period?
A project discussion records 7 days, followed by unrelated design decisions. StateTree recalls a caller-written note containing that same fact.
COMMIT NOTE
Export retention is 7 days. Completed export files are deleted after 7 days.
Authored fixture · no summarization model callTwo live inferences. Allow about a minute; a slow or offline PC can cause a timeout.
Ready when the operator’s model and tunnel are online.
02 / COMMIT CONTEXTWaiting
With StateTree
The question, selected note and StateTree overhead.
—input tokens
Output —Cached input —
MODEL ANSWER
The live answer will appear here.
Unknown usage and failed answers stay visible. A negative reduction means this example used more input tokens.
Inspect the complete report
Includes provider counts, task checks and selected commit IDs. This report is not a resumable agent session.
HOW THIS DEPLOYMENT RUNS
AWS hosts the app. Your PC runs Qwen.
- BrowserFixed comparison
- ECSStateTree + Strands
- SageMakerServerless gateway
- Local PCQwen via HTTPS tunnel
The operator’s PC, model process and tunnel must stay online. Completed cloud reports are stored privately in S3; an interrupted run does not resume after a worker restart. Report links act as unguessable access links.