netsi-claude-memory — "What has Claude actually remembered?"

Respond in Danish unless the user writes in another language (see the Language section below: all output is in Danish).

Claude Code automatically saves memories per project under ~/.claude/projects/<slugified-absolute-path>/memory/. The folder contains a MEMORY.md with one-line pointers plus one markdown file per memory with YAML frontmatter (name, description, metadata.type — user, feedback, project or reference) and a free-form body.

This skill opens that box and explains in clear Danish what is actually in it — not a raw dump, but something a person can act on.

Language

All output is in Danish. Even if the user writes in English, and even if the memory files are written in English. Quotes from the files are kept in their original language, but the framing around them is Danish.

The inviolable rule

Every single statement in the report must be traceable to concrete content in a concrete memory file. No filler, no qualified guesses, no "probably". If there are fewer than 5 genuine fun facts or fewer than 5 genuine need-to-know points, then say so explicitly ("Der er kun materiale til 3 fun facts i dette projekt") instead of padding the list.

---

Step 1 — list the projects, un-slugified

Run the helper script:

python3 scripts/memory_scan.py list

(If the script is not in the working directory, use the full path to the skill folder, e.g. ~/.claude/skills/netsi-claude-memory/scripts/memory_scan.py.)

It lists every folder under ~/.claude/projects/ and translates the slug back into a readable absolute path:

-Users-stenhougaard-Documents-GitHub-frh  →  /Users/stenhougaard/Documents/GitHub/frh

Be aware that un-slugifying is ambiguous: a - can be either a directory separator or a hyphen in a real directory name (sveltekit-netsi-dk is one segment, not three). The script solves this by walking the path on the filesystem and choosing the longest match that actually exists. If the path cannot be found again (the project has been deleted, or the folder comes from another machine), it is marked ⚠ gættet sti — pass that marker on to the user, do not hide it.

Present the result as a plain numbered list — path first, slug and memory count as secondary information. E.g.:

1. /Users/stenhougaard/Documents/GitHub/frh — 12 memories
2. /Users/stenhougaard/Documents/GitHub/netsi-dk — 4 memories (⚠ gættet sti)
3. /Users/stenhougaard/projekter/kunde-x — ingen memory-mappe

Step 2 — stop and ask

Ask the question and wait. Use AskUserQuestion if there are few enough projects for them to be options; otherwise ask in plain text: "Hvilket projekt vil du have gennemgået? Skriv nummeret."

Never proceed on your own — not even if there is only one obvious project, and not even if the user is standing in a project that matches one of the folders. The choice is the user's.

Exception: if the user has already named the project in their message ("gennemgå memories for frh"), still show the list, mark the match you found, and ask for confirmation of the number.

Step 3 — read everything

Once the number is chosen:

python3 scripts/memory_scan.py scan <nummer>

The script prints frontmatter, dates with source attribution, statistics and warnings — but not the body by default. Then read each memory file with the Read tool, so the summary is based on the actual content and not just on the description field. (Alternatively scan <nummer> --bodies, if you would rather have everything in one go.)

Handle these cases explicitly in the report rather than letting them disappear:

Situation	How to report it
No memory/ folder	"Projektet har ingen memory-mappe — Claude har intet gemt her." Stop.
Empty memory/ folder	"Memory-mappen findes, men er tom." Stop.
File mentioned in MEMORY.md but missing	Mention it under "Sprunget over" with the filename.
File cannot be parsed (invalid frontmatter)	Mention it under "Sprunget over" — and use the body anyway, if it is readable.
File on disk but not mentioned in MEMORY.md	Include it, and mark it as not indexed.

If there are no memories to analyse, say so plainly and do not produce the four sections below. An empty report is a valid result.

Step 4 — write the report

Four sections, in this order, with these headings:

"Om dette projekt"

A short summary in plain language of what each memory is about. Explain the meaning in context — not a repetition of the description field. Bad: "deploy.md: hvordan der deployes". Good: "Claude har lært, at sitet kører på Netlify med serverless-funktioner, og at adapter-static ligger i package.json uden at blive brugt — så en 'det er jo et statisk site'-antagelse vil være forkert."

One short bullet or one paragraph per memory file. Mention the filename so the summary can be verified.

"Fun facts"

At least 5 bullets. Specific, surprising or amusing details taken directly from the content. Each bullet must end with a source reference: (kilde: brugerprofil.md).

What counts as a fun fact: an unusual preference, an opinion the user has expressed sharply, a technical curiosity, a name or detail that surprises, a pattern in what Claude has been corrected on.

What does not count: generic project info ("projektet bruger TypeScript"), something you infer rather than read, or a rewording of a need-to-know.

"Need-to-know"

At least 5 bullets. The operationally most important things a new collaborator — or a future Claude session — must know before touching the project. Prioritise by what does harm if you don't know it: traps, inviolable conventions, things that have been corrected before, commands that must be run in a particular way. Here too: (kilde: <filnavn>) on each bullet.

"Metadata-analyse"

About the files themselves, not their content. Must contain:

• Ældste memory — date + filename.
• Antal pr. type (user / feedback / project / reference, plus any files without a type).
• Fordeling pr. ugedag and pr. måned — table or bullets.
• Mønstre der er værd at nævne: a cluster around one week, one type that dominates, a long stagnation, memories that were never indexed.

State the date source. The script reports per file whether the date came from a field in frontmatter (e.g. frontmatter.metadata.created) or from the file's mtime, and summarises the distribution. Write that source next to each figure, e.g.: "Ældste memory: 2. marts 2026 (brugerprofil.md, kilde: metadata.created)" and "Ugedags‑ fordelingen bygger på mtime for 2 af 4 filer, da de mangler datofelt."

Note to the user if many dates come from mtime: mtime changes when a file is edited or copied, so the distribution then says more about last modified than about when the memory came into being.

All dates are converted to the machine's local time before being distributed by weekday and month. A frontmatter field may carry a time zone (2026-03-02T23:30:00+02:00), while mtime is always local — without the conversion the two would sit on different timelines, and a time close to midnight would land on the wrong weekday.

---

Helper script

scripts/memory_scan.py (Python 3.9+, no required dependencies — uses PyYAML if it is installed, otherwise a built-in minimal parser).

python3 scripts/memory_scan.py list                    # nummereret projektliste
python3 scripts/memory_scan.py scan 3                  # læs projekt nr. 3
python3 scripts/memory_scan.py scan 3 --bodies         # tag brødteksten med
python3 scripts/memory_scan.py --json scan 3           # rå JSON
python3 scripts/memory_scan.py --root /sti/til/rod list  # anden rod-mappe

A directory name can be used directly instead of the number (scan -Users-navn-projekt) — the script itself handles the name starting with -.

scripts/test_memory_scan.py is a regression test for the script (run python3 scripts/test_memory_scan.py; exit 0 = everything passed). Among other things it covers the date conversion across time zones and the extraction of pointers from MEMORY.md — update it too if you change those parts.

If you do not have the script at hand, the steps can also be run by hand with Glob, Bash (ls ~/.claude/projects/) and Read — but then the un-slugifying, the date sources and the statistics must be done manually and with the same care.