Solution
Your context window fills and the agent forgets the morning. Keep working in one session.
Context Mode keeps older work word for word in an archive and sends the model recent work plus an index. In a paired run on a 200K window: no compactions, 9 of 10 early details exact, 37.6% lower cost.
The same 30 steps on a 200K window, run twice.
37.6% lower per session. Cheaper in 3 of 3 pairs, Claude Opus 5.5.
How often the session compacted.
Through Context Mode the session never compacted.
After the work, 10 questions about early details.
Exact answers. Task checks 8 of 8 in both arms.
Early on you said: keep the refund window at 14 days.
After compaction
- In the summary only if the summary kept it
- In other words, not yours
- The agent may ask again
With Recall
- In the index, word for word
- Listed as a decision the agent must not reopen
An example of the mechanism, from the Recall docs.
What you see when Recall acts.
⚡ context-mode · picked up after a 2 h pause, nothing compacted
⚡ context-mode · recalled 2 items from 2 days ago (a decision, src/billing/quota.ts)Each line shows once. The console lists each lookup.
The window sets your day. Not your work.
Close to the window, your agent compacts: it swaps older history for a summary. From then on it works from the summary, not from your morning.
The summary drops exact lines and early rules. The agent asks the same question again, or makes the other choice.
Until it compacts, every request sends the whole history again. Each request costs more than the one before.
Set a limit once. Work all day.
- 01Pick a size
Choose 200K, 500K or 1M in Active Context. Each thread stays under it.
- 02Older work goes to the archive
Word for word, before it leaves the model's view. The model gets recent work and an index of the rest.
- 03The agent looks it up
It fetches an early detail by id, or finds it by search. You do not type it again.
Want to see it step by step? One session, from hour one to hour six →
The same workday, run twice.
3 pre-registered pairs on 2026-10-04, mean of 3 pairs. Claude Code on Claude Opus 5.5, the same 30 steps in both arms. Cheaper in 3 of 3 pairs. Task checks 8 of 8 in both. Method and sources →
What this uses.
Recent work and an index go to the model. Older work stays in the archive, word for word.
Pick 200K, 500K or 1M once. Each thread stays under it.
Find any prompt, command or tool output from past sessions, in the console or from your agent.
Related: All-day session, step by step · All-day sessions without compaction · Session resume · Recall docs
One session, all day.
npx @context-mode/cli connects your agents. 1,000 requests free, no card.