Context lifecycle
Why every turn costs more than the last
Agents never clean up after themselves. Here is what that costs, and what Vachi does about it.
You are paying for stale context
Your coding session accumulates stale context
Old errors, superseded plans the context keeps piling up. Your session moves on. But the stale context remains.
You pay for it over and over
Claude Code resends the whole session context with every request. Even with prompt caching you lose over half your usage to stale information.
Agents never clean up stale context.
If processes handled memory the same way.
Your OS manages memory lifecycle
to maximize device performance
foreground app1.8 GB
open document340 MB
kernel + drivers900 MB
background tabinactive1.2 GB
minimized appidle640 MB
installer cachefinished2.1 GB
Vachi manages context lifecycle
to optimize cost per task for agents
system prompt3.2k
src/auth/session.ts9.8k
diff · refresh token on expiry5.2k
npm test · 3 failingresolved3.4k
plan v1 · patch validatorsuperseded3.6k
docs · session api referencenot needed now13.5k
Useful information is preserved in context. Paginated information is recalled on demand.
Vachi preserves quality
We validated that output quality remains unchanged across several benchmarks.
Removing stale context mitigates bad outcomes
Driven by improved signal to noise ratio.
- Vachi AISWE-Session
Same model. Same quality. Less to resend.