Should I use CLAUDE.md? Answer four questions.

Your answers

Full pass for each kind of memory

Each dot is the share of sessions with a full pass. Each bar is the 95% interval of that share. Point at a row to see which intervals overlap it.

What the study measured for your answers

Every number shows its n. A rate shows its 95% interval. A calculation says so.

• Facts the code does not show

  Calculation

  • A team fact is one that no file in the repository holds, such as the changelog rule or the late-fee rate.
  • With no memory, Claude Sonnet 5.5 passed 6/15 team-knowledge checks (95% interval 19.8% to 64.3%, n = 15).
  • With the curated 11-line file, it passed 15/15 (95% interval 79.6% to 100%, n = 15).
  • The two intervals do not overlap (calculation).

  • 6/15 Team-knowledge checks, no memory

    95% interval 19.8% to 64.3%, n = 15

  • 15/15 Team-knowledge checks, curated 11-line file

    95% interval 79.6% to 100%, n = 15

• The late-fee rate

  • A task told Claude Sonnet 5.5 to charge "our standard late fee". No file of the repository held the rate.
  • When any memory file held the rate, Claude Sonnet 5.5 used it in 15/15 sessions (95% interval 79.6% to 100%, n = 15).
  • With the raw notes, which also held an old rate, it used the current rate in 3/3 sessions (n = 3, no interval stored).
  • With the rate in no memory file, it stopped to ask in 7/9 sessions (n = 9, no interval stored) and wrote no code.
  • The other 2 guessed a rate and said it was a placeholder.

  • 15/15 Late fee right when any memory file held the rate

    95% interval 79.6% to 100%, n = 15

  • 3/3 Late-fee sessions that used the current rate, raw notes

    n = 3, no interval stored

  • 7/9 Without the rate: sessions that asked and wrote no code

    n = 9, no interval stored

What every answer shares

• Full pass with Sonnet 5.5

  Calculation

  • A full pass means that the hidden tests and every convention check pass.
  • With no memory, Claude Sonnet 5.5 reached a full pass in 9/15 sessions (95% interval 35.8% to 80.2%, n = 15).
  • With the curated 11-line file, it reached a full pass in 15/15 sessions (95% interval 79.6% to 100%, n = 15).
  • The two intervals overlap, so the study does not separate them (calculation).

  • 9/15 Full pass, no memory

    95% interval 35.8% to 80.2%, n = 15

  • 15/15 Full pass, curated 11-line file

    95% interval 79.6% to 100%, n = 15

• The /init file and a stale README command

  • The README held a test command that fails on Node 25. The /init file copied it.
  • Claude Sonnet 5.5 ran that command in 13/15 sessions with the /init file (95% interval 62.1% to 96.3%, n = 15).
  • With no memory, it ran the command in 12/15 sessions (95% interval 54.8% to 93.0%, n = 15).
  • With the curated 11-line file, it ran the command in 0/15 sessions (95% interval 0% to 20.4%, n = 15).

  • 13/15 Stale test command, /init file

    95% interval 62.1% to 96.3%, n = 15

  • 12/15 Stale test command, no memory

    95% interval 54.8% to 93.0%, n = 15

  • 0/15 Stale test command, curated 11-line file

    95% interval 0% to 20.4%, n = 15

• What memory costs in context and time

  • The median input counts every turn of a session, cache reads included.
  • For Claude Sonnet 5.5, the median input was 78,455 tokens with no memory, and 65,045 with the curated 11-line file.
  • The 210-line handbook gave 85,223 tokens, and the Stop hook alone gave 125,674.
  • The hook adds turns: each block sends the agent back to work.
  • Each value has n = 15. The dataset stores no interval for a median.
  • The median time per session was 18 s with no memory and 21.9 s with the curated 11-line file (n = 15 each).
  • Up to four sessions ran at a time on one machine.

  • 78,455 Median input tokens, no memory

    n = 15, median, no interval stored

  • 65,045 Median input tokens, curated 11-line file

    n = 15, median, no interval stored

  • 85,223 Median input tokens, 210-line handbook

    n = 15, median, no interval stored

  • 125,674 Median input tokens, Stop hook alone

    n = 15, median, no interval stored

  • 18 s Median time, no memory

    n = 15, median, no interval stored

  • 21.9 s Median time, curated 11-line file

    n = 15, median, no interval stored

• Cost per fully correct result

  Calculation

  • The study adds the list-price cost estimates of a kind of memory and divides the sum by its full passes (calculation).
  • For Claude Sonnet 5.5, the result was $0.14 with no memory (9 full passes) and $0.082 with the curated 11-line file (15 full passes).
  • A failed session still costs money, so cost per correct result falls when fewer sessions fail.
  • The sessions ran on a subscription. These values are list-price estimates, not invoices.

  • $0.14 Cost per full pass, no memory

    n = 9, n counts full passes, no interval stored

  • $0.082 Cost per full pass, curated 11-line file

    n = 15, n counts full passes, no interval stored

What memory costs

The same kinds of memory for the same model. The study measured the tokens. The cost is a calculation at list prices.

How your answers pick the highlights

Each answer highlights fixed kinds of memory. The map reads what a kind holds. It does not read the results.

Yes, your team knows such facts

Curated, 11 lines, Raw notes, 60 lines, Dreamed notes, Handbook, 210 lines, Curated + hook

Yes, you need rules enforced

Stop hook only, Curated + hook

Yes, your notes grow without review

Dreamed notes

A "no" answer highlights nothing.

What could make this picker misleading

The study states its own limits. The picker adds two more.

• One small synthetic repository and one author of the facts: the curated file is an upper bound written with knowledge of the tasks.
• The hook checks the same code rules as the grader. It shows what rules written as code can do; it cannot carry a fact such as the late-fee rate.
• n = 15 per condition for Sonnet and 10 for Haiku: most full-pass intervals overlap, so most differences between conditions are not clear.
• Costs are the CLI's list-price estimates for subscription sessions, not invoices.
• In headless mode an agent that asks a question cannot get an answer, so asking counts as a failure here. In a live session the person would answer; the cost is the round trip.
• The picker highlights kinds of memory. It never scores or ranks them.
• Your repository is not this repository. Use the picker to choose what to test, not to skip the test.

Questions about CLAUDE.md

Read the evidence

The study: 8 kinds of agent memory, tested The post: does CLAUDE.md help? What is CLAUDE.md? The Stop hook, tested Memory consolidation (dreaming), measured Claude Haiku 4.5 and Claude Sonnet 5.5, compared Claude Sonnet 5.5: every measured value