Capture one durable fact as a recipe card under ~/.typmem/memory/learnings/. A learning is what you now know about the world — useful to anyone, indefinitely, even someone who never saw the painful session that produced it. A learning is valid and usable the moment you capture it — it does not wait for any review. typmem-judge later tidies learnings and may propose that the strongest ones be promoted to precedents/ or rules.md, but that is optional and non-blocking.
Learning vs retro — read this first
- Learning = recipe card. "What I now know about the world." A distilled fact. No story, no timeline, no
I. True regardless of which day or project hit it. → goes here. - Retro = diary. "What happened to me." Status, decisions-in-flight, dead ends, follow-ups. → goes to
typmem-retro.
The test: "Will this still matter after the work that produced it is finished?"
- Yes → it is a learning. Capture it here.
- No → it is retro material. Do NOT write it here; it belongs in
typmem-retro.
What makes a good learning
- Atomic — one fact per file, so each can be searched, judged, and promoted on its own. If you have three facts, write three files.
- No
I, no timeline, no story. State the fact, not the journey that produced it. - Self-contained — understandable by someone who never saw the work behind it.
- Specific and actionable — a constraint, a gotcha, a pattern, a how-things-work truth. Not a vague impression.
Example of the move — turn a diary sentence into a recipe card:
Diary (retro): "Spent two hours today fighting the iii-engine install, the binary kept failing on this machine, eventually gave up."
Recipe card (learning): "iii-engine has no prebuilt binary for this arch — anything depending on it needs the Docker fallback."
Two ways to use this
- Direct capture (default). You (or the user) have a fact in mind right now → write one learning file.
- Extract from a retro. Point at a retro file (or the latest one in
~/.typmem/memory/retro/) → read it, pull out each durable fact, and write one learning file per fact. Distil — do not copy the narrative across.
How to write it
- Make sure the directory exists:
mkdir -p ~/.typmem/memory/learnings - Distil the fact and run the test above. If it fails (it only matters for this session), stop — it belongs in
typmem-retro, not here. - Build a slug from the fact itself (not the session): lowercase kebab-case, 2–6 words, no spaces (e.g.
iii-engine-no-prebuilt-binary,aca-sidecar-no-host-network). - Target file:
~/.typmem/memory/learnings/<slug>.md— no date prefix. Learnings are timeless; only retro carries a date in the filename. - Write this frontmatter, then the body:
Leave--- author: <agent name or person capturing this> reviewer: created: YYYY-MM-DD ---reviewerempty. An emptyreviewerjust meanstypmem-judgehas not looked at this one yet — it does not mean the learning is unusable. The learning is valid right away. Do not fillreviewerin from this skill. - Body: state the fact in recipe-card shape — a short, self-contained claim. Free-form prose is fine; no fixed template yet. Keep the discipline: no
I, no timeline. - If a file with that slug already exists: it is probably the same fact. Either refine the slug to be more specific, or update the existing file rather than creating a near-duplicate. (Automatic dedup comes later with
typmem-judge/ index search — for now, use judgement.)
When extracting from a retro, repeat steps 2–7 once per durable fact you find.
File rules
- Location:
~/.typmem/memory/learnings/ - Filename:
<slug>.md— no date prefix (timeless fact) - Frontmatter (MVP):
author,reviewer(empty —typmem-judgefills it once it has looked),created - One fact per file
Don't
- Don't write session narrative, status, or what-happened here — that is a retro (
typmem-retro). - Don't fill in
revieweryourself — that istypmem-judge's job; an empty field just means the judge hasn't looked at it yet, not that the learning is unusable. - Don't bundle multiple facts into one file — split them so each can be judged and promoted independently.
- Don't promote a learning to
precedents/from here — that is a human review step handled by a later skill.