Promote daily notes into agent knowledge that compounds

1. Tag during capture, promote during review

The split is surgical. Capture is fast, promotion is deliberate. Never try to do both at once.

During the day, in your daily note, tag moments worth promoting with an inline marker, #learning, #pattern, #decision, whatever your vault convention is. The tag is a promise you'll look at this again on Sunday, not a finished thought.

Do not write the lesson while the thing is still happening. You can't judge importance from inside the moment. The weekly review is the promotion step, on purpose.

2. Sweep the week's tags in one pass

Sunday (or whichever day becomes your review slot), open every daily note from the week. Grep for your tags. Pull each tagged moment into a list.

You should end the sweep with somewhere between zero and ten candidates. Less than that means you're under-capturing. More than ten and you're promoting noise, raise your capture bar mid-week, not at review time.

3. Write each promotion in the three-question format

For every candidate that survives the sweep, write a short file under 06-reference/learnings/YYYY-MM-DD-<slug>.md. Every learning answers three questions:

• What happened, one short paragraph, specific. File paths, decisions, outcomes. No generalizing yet.
• Why it mattered, the mechanism. Why did this outcome occur? What's the underlying rule?
• When this applies, the future-case predicate. What situation should make you reach for this lesson again? Be precise.

Keep each learning under 200 words. If it wants to be longer, it's probably two learnings.

4. Link every learning back to the work that triggered it

Cross-linking is the difference between a pile of notes and a living wiki. Every learning links back to:

• The daily note entry that triggered it
• The project or file it applies to
• Any related prior learnings (this is where compounding happens, two learnings from different months connecting becomes a pattern)

Obsidian's [[wikilinks]] syntax works, plain paths work, it doesn't matter, just make the connection explicit. Agents follow the links. So do you, six months later, when you've forgotten the context.

5. Promote up, archive down

Every few months, do a second-pass sweep. Re-read the learnings from three months ago. Decide:

• Promote up, this learning shows up repeatedly across unrelated work. It deserves to live in soul.md or a higher-level reference doc as a durable principle.
• Archive down, this learning was real for its time, but the situation has passed. Move to 07-archive/ so it stays grep-able without cluttering the active reference layer.

Without this second pass, 06-reference/ bloats. The goal is to keep it small and dense. A tight reference layer is an agent feature.

6. Verify the loop at the three-month mark

Three months in, test the compounding. Open a brand-new agent conversation with no prior context. Ask it: "Based on what you know about me from this vault, what patterns do I keep repeating?"

If it surfaces patterns you recognize, the ones you thought were insights three months ago, the loop is working. If it gives you generic fluff, your promoted learnings are too abstract. Rewrite three of them more specifically and try again next month.

---

Why this rhythm beats "I'll remember" and "I'll organize later"

Promotion under pressure is broken promotion. Writing the learning while the project is still hot means you over-emphasize what just happened. The weekly delay is a feature, it filters out noise that felt big on Wednesday.

Search finds what you promoted, not what you captured. Your agent's usefulness scales with the size and quality of the promoted layer, not the raw capture layer. Capture is a receipt. Promotion is the asset.

Archived is better than deleted. Grep still finds archived notes. You keep the semantic trail. Nothing is lost, but the active reference layer stays small. This is the only way it scales past a thousand files.

Steal this starter

Learning template (06-reference/learnings/YYYY-MM-DD-<slug>.md):

---
date: YYYY-MM-DD
tags: [learning, <topic>]
source: [[01-daily/YYYY-MM/YYYY-MM-DD]]
related: [[06-reference/learnings/<prior-slug>]]
---

# <one-line title that names the pattern, not the event>

## What happened
<one short paragraph, specific, file paths, decisions, outcomes>

## Why it mattered
<the mechanism, underlying rule, not restatement>

## When this applies
<the future-case predicate, what situation triggers this lesson>

Weekly review checklist:

• Open every daily note from the week
• Grep for #learning, #pattern, #decision (or whatever tags you use)
• Pull each hit into a shortlist
• For each survivor: write the three-question file
• Cross-link: source daily note, affected project, related prior learning
• Commit the promoted files (git add 06-reference/ && git commit)
• Delete the tag from the daily note so the sweep stays clean next week

Grep one-liner for the sweep:

grep -rn --include="*.md" -E "#learning|#pattern|#decision" 01-daily/$(date -v-7d +%Y-%m) 01-daily/$(date +%Y-%m)

That's the whole rhythm. Thirty minutes on Sunday. After six months, the 06-reference/learnings/ folder is the single most useful context you hand to every agent, every time.