← The AI Hype Audit — all 645 verdicts

REAL

Holds up: a short 'map' file that loads first, pointing to plain-text notes, people, project and lessons files, is a real and sound way to give Claude Code a working memory, and it is the same shape Claude Code's own auto memory uses

Facebook · Nov 2026

The claim'One of the most underrated moves in Claude Code': one map file that spins out into plain-text files. Notes (one topic per file), people (a card per person), projects (status and next step), and a root folder for long-term memory, lessons and a decision log. The map loads before you ask anything, and every answer ends either sourced (one file opened and named) or flagged as never stored.

This is a good, honest description of how to give Claude Code a memory, and it matches Anthropic's own design. Claude Code's documentation says CLAUDE.md files in your project and user folders load at the start of every session, and that its built-in auto memory works exactly this way: a short MEMORY.md index (first 200 lines or 25KB load at startup) that points to separate topic files, which Claude opens on demand with its normal file tools. So 'the map loads before you ask anything' is true for the map file itself; the notes, people and project files do not load until Claude decides to read them, which is the point, because it keeps the context small. One caveat: Anthropic is explicit that these files are context, not enforced rules. 'Every answer ends sourced or says it was never stored' is an instruction you write into the map, and the model will follow it most of the time, not every time. If you need a hard guarantee, the docs point to hooks. Also note the anti-pattern: using @imports for every file loads them all at once and defeats the purpose. Only the post text was captured; any carousel slides were not, so this grades the written claim. Nothing is being sold in the captured text; it is a community tip in a user group.

What holds up

  • Read Claude Code's memory documentation (code.claude.com/docs/en/memory): CLAUDE.md in the directory hierarchy loads at launch; subfolder files load on demand.
  • Docs confirm auto memory uses a MEMORY.md index (first 200 lines or 25KB loaded) plus topic files that are not loaded at startup and are read on demand.
  • Docs state memory is 'context, not enforced configuration'; hard rules need a PreToolUse hook.
  • Docs warn @imports load with the importing file and do not reduce context cost. x-ref #435, #545.

What doesn’t

  • 'Every answer ends sourced' is a prompt rule, not a guarantee.
  • Only the map loads up front; the other files are read only when Claude chooses to open them.
  • Carousel slides were not captured; graded on the post text only.

The catch

The pattern is right, but the map file is advice to the model, not a lock. Claude follows a well-written index most of the time; it can still skip a file or answer from its own head.

How to actually do it

  • Create a CLAUDE.md under 200 lines that lists each notes, people and projects file with one line on what it holds and when to open it.
  • Keep detail in the separate plain-text files and link them by path in backticks, not @imports, so they load only when needed.
  • Add one rule to the map: 'cite the file you used, or say the fact is not stored', and spot-check answers weekly.
  • If a rule must never be broken, enforce it with a Claude Code hook instead of a sentence in the map.

A free community tip that describes a real, documented Claude Code pattern. The only oversell is implying the sourcing rule always fires.

Confidence
High
Posted by
a member post in a Claude Cowork Facebook group; post sent to Buddy Fri Nov 13, 2026, 6:33 AM

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.

© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.