Product · Context management & collaboration

Context Manager

From scattered signals to coherent intelligence.

Long sessions lose their thread when the conversation is compacted or the model changes, and you pay to rebuild it. The Context Manager keeps a short working memory of the session, so the next model picks up where you left off instead of starting again.

Start free trialIn Cursor, Claude Code and Codex, on your machine.

98%

of calls re-read the whole previous conversation, and it is billed every time.

59%

of what gets re-read is the model's own earlier output and thinking.

$666 → $274

the same sessions, replayed with a fresh start at the right moment: 59% less.

In our logs: 4,393 Claude Code calls from one heavy user, October 2026, priced at list rates. They measure the problem long sessions have; they are not a measured OptiFlume saving.

What the next model is told.

One short note, kept as you work. Not a blurry summary of everything: the four things a model needs to carry on.

01

What are we trying to do?

The goal of the session, in a sentence or two, so the next model works toward the same thing.

02

What has been decided?

The choices already made along the way, so they are not reopened.

03

What must not change?

The constraints you set: the library to keep, the file not to touch, the style to follow.

04

What already failed?

Approaches that were tried and did not work, so the next model does not try them again.

Before, and with OptiFlume.

  • BeforeAfter compaction the model forgets what you agreed.

    WithGoals, decisions and constraints come through in a short note.

  • BeforeRe-explaining the task after every compaction.

    WithThe model picks up where you left off.

  • BeforeThe failed fix, tried for the third time.

    WithFailed approaches are on the note, so it moves on.

  • BeforeA new model mid-session, starting from a summary.

    WithAfter a compaction, a model OptiFlume switches to gets the same note.

01

Working memory

When your tool compacts a long conversation, the session's goals, the decisions made and the constraints you set are handed to the model in one short note. A model OptiFlume switches to later in the session gets it too, not only your tool's summary.

02

Failure memory

Approaches that already failed are carried in the same note, so the next model does not try them again. Fewer loops, fewer wasted turns.

03

Across your tools

Coming next: start a session in one tool and continue it in another. Work begun in Cursor will carry into Claude Code or Codex without explaining it all again.

04

Yours, on your machine

Working memory is kept under ~/.optiflume on your machine, in files you can open and delete. It is never sent anywhere. On Pro and during the trial.

The same in every tool.

In Cursor, Claude Code and Codex, in VS Code and the terminal: nothing to click and nothing to configure. It keeps working while OptiFlume is paused.

  • Cursor
  • Claude Code
  • Codex

Questions.

In the docs: Working memory and session watchdog · What is stored where

All questions →
Do I have to do anything?+

No. There is nothing to click and nothing to configure. The note is kept as you work and handed over after your tool compacts a long conversation.

Where is it kept, and can I delete it?+

On your machine only, under ~/.optiflume, with the cached turn output it is built from kept for 30 days. Neither is ever sent anywhere, and you can delete both whenever you like.

Can I continue a session in another tool?+

Not yet. Carrying one session's context from one tool to another is coming next.

Is it on the free tier?+

Working memory is part of Pro and the 30-day trial. The session watchdog, which says when a session is going in circles, is free on every plan.

Start with your next prompt.

Pro is free for 30 days, no card. Install, keep working, and see what it did and why.

macOS 11+ · Apple silicon and Intel · Signed by OptiFlume, notarized by Apple · Never locks you out · Uninstall with one command