Skip to main content
mnemom agent shows what Mnemom saves you in two places:
  • a first-run savings screen, a one-time look back at your recent Claude Code sessions;
  • a savings counter in Claude Code’s status bar while a session runs.
Both are estimates, worked out on your machine. Long sessions are where Mnemom saves the most: from about 20% to over 35% in our datasets, compared with stock Claude Code.
These figures come from Mnemom datasets chain45 and chain240, with Claude 5.5-series models.

First-run savings screen

The first time you launch mnemom agent in a terminal, it asks before Claude Code starts:
To see the screen again at any time, run mnemom agent --first-run. It shows the screen and exits without launching.

What it shows

The screen covers your last 30 days of Claude Code on this machine:
  • the tokens you used and what they cost at Anthropic API list prices. Everything is priced at API prices, whatever you actually pay, including on a subscription;
  • about how many fewer tokens you would have used with Mnemom, as a number and a percentage, and about how much less it would have cost;
  • your biggest session, and about how much less it would have cost with Mnemom;
  • for sessions that already ran through Mnemom, their usage and what Mnemom saved on them, shown apart from the rest and what Mnemom would have saved there;
  • on a Claude Team or Enterprise plan, when a saving is shown, a line saying that fewer tokens means more work before you hit a limit (and less overage, if extra usage is on). It never shows a dollar figure for your plan.
The screen shows a saving only if it comes to at least $5 and at least 5%. If no saving reaches that, the screen says your sessions stayed small, so Mnemom wouldn’t have changed much. If there is too little history to look at, it says so. In both cases it shows the general figure for long sessions instead:
That figure is on Mnemom datasets chain45 and chain240, with Claude 5.5-series models. Every number on the screen is rounded down.

What it reads and writes

  • It reads Claude Code’s session files on this machine, and only their usage data: token counts, model names, timestamps, and Claude Code’s own compaction and usage-limit records. It never reads the content of your messages.
  • It makes no network calls. Nothing is uploaded.
  • It writes two files in ~/.mnemom, readable only by you: first-run.json remembers your answer, and first-run-audit.json records each number on the screen and how it was worked out.
You are never asked in a script, a pipe or CI, or with --dry-run. To never be asked on a machine, set MNEMOM_AGENT_NO_FIRST_RUN=1.

Savings counter

When mnemom agent launches Claude Code, it adds a segment to Claude Code’s status bar that updates as the session runs:
If you already have a status line in your Claude Code settings, it keeps running and the counter is added after it, separated by ·. The counter uses the same estimate as the first-run screen, applied to the current session, including its subagents, and rounded down to the cent. It reads the session’s files on this machine and makes no network calls. It keeps a small state file per session in ~/.mnemom/statusline/.

Turn the counter off

Set MNEMOM_AGENT_NO_STATUSLINE=1 before launching:
There is no mnemom agent config setting for it. The counter is also left out when you pass your own --settings to Claude Code after --, and it is only added when the coding-agent CLI is Claude Code.

How the estimate works

  • Each request in a Claude Code session file is priced at Anthropic API list prices for its model, from the token counts Claude Code records.
  • Mnemom saves by folding long context. Requests too small for Mnemom to fold count as no saving.
  • Larger requests get a saving rate from our benchmarks of long Claude Code sessions run with and without Mnemom: one rate for mid-size requests, and a higher one for very large ones. The number of requests and the output are assumed to be the same as without Mnemom.
  • A rate per size band overstates the saving on some requests and understates it on others. Across many sessions it evens out.
  • For a session that already ran through Mnemom, the first-run screen works the other way round. It replays the session as stock Claude Code would have run it and counts the difference. That saving can be negative for a single session.
The estimate is not your bill. Prices, plan limits and your own sessions vary, and some background calls Claude Code makes are not in its session files.