compile-knowledge

Durable knowledge is worth keeping as many small, interlinked markdown files, compiled over time and surfaced through an index — not as one giant doc, a chat log, or a one-off file that rots. This skill makes compiling consistent so your agent gets smarter over time instead of relearning the same things.

Where it goes — pick the right store

• Agent memory (default, always available): your .claude/.../memory/

folder with MEMORY.md as the index. This is the per-agent store and it survives restarts — it's the karpathy "external memory" that keeps you sharp across sessions. Governed by the memory rules already in your system prompt — follow them. For most agents this is the only store you need.

• Shared wiki (only if you work as a team): a wiki/ folder in your project

with a wiki/index.md. For knowledge the whole team benefits from — domain facts, research findings, reference material multiple agents would re-derive. Skip this entirely if you're a solo agent; don't manufacture team ceremony.

Rule of thumb: "only I act on this" → memory. "Anyone on my team might need this" → shared wiki. Cross-link between them with [[slug]] when they relate.

The async pipeline vs. you — division of labor

Some platforms run an automatic consolidation pass over finished sessions (on 5dive: 5dive memory consolidate, scheduled for you by the heartbeat). If your platform has one, know what it covers and what only you can do:

• AUTOMATIC — plain facts. The pass distils your FINISHED session

transcripts into memory atoms in your own store. It never reads the live session, and nothing it writes leaves your box. You do NOT need to hand-copy plain facts out of a session to keep them — that is what stops knowledge dying with the context window.

• STILL YOURS — judgement. A wiki page, a decision and its reason, a gap

analysis, the CAUSE behind a finding: the pipeline can only lift what is stated in the transcript, and a conclusion you drew is not lying there. It also cannot publish to a shared wiki — it has no store selector, and shared stores are deny-default by design — so publishing to the team store stays a curated act, and that act is yours.

This narrows what you compile; it does not replace compiling. Judgement-shaped knowledge still goes through the procedure below, before you close the task.

Before you write — the hygiene gate

Compile ONLY a durable, non-obvious fact. Skip and move on if it is:

• routine / derivable from the repo, git history, or existing docs,
• true only for this one conversation,
• already covered by an existing file (→ UPDATE that file instead, don't duplicate).

Most tasks (a deploy, a restart, a one-line fix) produce nothing durable. That is fine — do not manufacture a memory to "have written something." Filler is worse than nothing; it pollutes recall.

The procedure

1. Search first. Look for an existing file on this topic (grep the store +

skim the index). If one exists, edit it — never create a near-duplicate.

1. Atomic. One fact / one topic per file. If you're tempted to add a second

unrelated fact, that's a second file.

1. Name it. kebab-case slug, with a type prefix for memory

(feedback…, project…, reference…, user…) or a clear topic slug for the wiki. The slug is the link target.

1. Frontmatter. name (= the slug), description (ONE line — this is what

gets matched during recall, make it specific), and a type/category.

1. Body. State the fact plainly. Link related entries with [[slug]] —

liberally; a link to a file that doesn't exist yet is a fine TODO marker. For feedback/project, follow with Why: and How to apply: lines.

1. Index. Add or update a ONE-LINE pointer in the index (MEMORY.md for

memory; wiki/index.md for the wiki): - [Title](slug.md) — hook. Keep it under ~200 chars; detail lives in the file, never the index. If the wiki index.md doesn't exist yet, create it so the store stays discoverable.

1. Hygiene. Delete files that turned out wrong. Convert relative dates to

absolute. If the index is getting long, tighten lines — don't let it bloat.

Lifecycle — supersede, don't silently overwrite

Facts rot. When one is time-sensitive, uncertain, or replaces an older one, stamp the envelope so recall can age it out instead of surfacing stale truth:

• valid_to: YYYY-MM-DD — date the fact expires / needs a recheck. Past it,

recall keeps the fact but demotes + flags it ⚠ expired.

• supersedes: <slug> — when a new fact replaces an old one, point at the

old slug. The old fact is then demoted (⤴ superseded) in recall instead of lingering as a second, contradictory answer. Prefer this over edit-in-place when the OLD value is still worth seeing (audit trail); edit in place when it isn't.

• confidence: high|medium|low — low/medium facts are demoted so a hunch

never outranks a verified fact.

• provenance: "<source>" — where the fact came from, distinct from

compiled_by (who wrote the note).

If your store has a CLI, it likely exposes these as flags — on 5dive, 5dive memory add takes --valid-to=, --supersedes=, --confidence=, --provenance=. All optional — omit them and behaviour is unchanged, and the frontmatter fields above are the portable part.

Anti-patterns

• A wall-of-text doc instead of atomic files.
• Research left as a standalone notes.md that never gets folded in — that's

working notes, not knowledge. Compile the durable parts into the store.

• Duplicating a fact across memory AND wiki — pick one home, cross-link.
• Index entries that restate the whole file.
• Writing filler to satisfy a habit/checklist.

Quick checklist

[ ] durable & non-obvious? [ ] right store? [ ] updated existing vs new? [ ] atomic + named + frontmatter? [ ] [[links]]? [ ] index line added?