dream

name: dream description: Consolidate recent session signal into your durable storage — memory atoms, mentality, wiki. Run on a cadence (typically daily, woken by pulse) or when you've just learned something worth keeping. The four-phase distillation Karpathy / claudefa.st AutoDream: orient → gather → consolidate → prune. Cheaper than re-reading transcripts — only matched lines + surrounding context get attention. allowed-tools: [bash, read, write, edit]

dream — distill recent sessions into durable storage

You wake into this skill when something tells you to "dream", "run dream", "consolidate", or "do your morning reflection". The pulse-fired daily wake-ups for most cells route here. The point: every so often, pull the durable signal out of recent sessions into the storage that survives context resets — state/memory/, state/mentality.md, state/wiki/.

The work is four phases. Be surgical. Touch what matters; skip the rest.

1. Orient — what storage do you have?

Check which storage exists. Only act on what's present.

ls state/memory/MEMORY.md state/mentality.md state/wiki/index.md 2>&1 | grep -v 'No such'

Then read the index of each that exists:

• state/memory/MEMORY.md — the index of memory atoms. Read it.
• state/mentality.md — current mentality. Read it.
• state/wiki/index.md — wiki page index. Read it.

If none exist, this cell has no durable storage yet. Stop here and note that to the requester.

2. Gather — what signal happened since the last dream?

Read the cursor (last successful dream timestamp). If no cursor, default to "24 hours ago".

cat state/.dream/cursor 2>/dev/null || date -u -d '24 hours ago' '+%Y-%m-%dT%H:%M:%SZ' 2>/dev/null || date -u -v-24H '+%Y-%m-%dT%H:%M:%SZ'

Find your session files. The path depends on your harness — pick the one that exists:

List files modified since the cursor. Drop the freshest — it's probably the active session and you don't want to digest yourself mid-thought.

find ~/.claude/projects/-root -name '*.jsonl' -newermt "$CURSOR" | sort | head -n -1

For each file, grep for the patterns below — don't read whole transcripts, that blows your context.

grep -niE '\b(actually|wait,|let me correct|i was wrong|remember (this|that)|save this|note that|important:|FYI|TIL|we decided|let'\''s go with|going to|the plan is|i (prefer|like|hate|don'\''t like)|always|never|used to think|changed my mind|turns out)\b' <session.jsonl>

Pull each matched line plus ~2 lines of context. These are candidates, not signal. The next phase filters them.

If grep returns nothing across all files, the cell had no durable signal since the cursor. Skip phase 3, write the cursor, log a no-op entry to wiki/log.md, and report back "no signal".

3. Consolidate — write findings into storage

Filter the candidates for what's actually durable:

• User corrections ("actually X, not Y") → memory atom or mentality "mind changes"
• Explicit saves ("remember that…", "important:") → memory atom
• Key decisions ("we decided…", "going to…") → memory atom or mentality update
• Recurring patterns (same topic across multiple sessions) → wiki page
• Things you used to think but no longer → mentality "mind changes"

Skip:

• Tool-call mechanics ("I'll run bash to…")
• Conversational acknowledgments ("got it", "ok")
• Anything already saved (cross-check against orientation snapshot)

Writing memory atoms (state/memory/)

Naming convention (the memory package validates these):

• user_<topic>.md — facts about the user
• feedback_<topic>.md — user corrections / preferences
• project_<topic>.md — ongoing work
• reference_<topic>.md — pointers to external systems

Each atom is small (one fact, a few lines). Frontmatter:

---
name: <kebab-case-slug>
description: One-line hook
metadata:
  type: <user|feedback|project|reference>
---

The fact. For feedback/project: lead with the rule/fact, then **Why:**
and **How to apply:** lines.

If an atom on the same topic exists, edit it (merge, refresh) — don't create a duplicate. Then refresh its line in state/memory/MEMORY.md (keep that index ≤200 lines).

Updating mentality (state/mentality.md)

Augment, don't rewrite. Usually a "Mind changes" or "Lessons learned" entry. Keep total file ≤80 lines / 6KB.

Writing wiki pages (state/wiki/)

Genuinely new topic from a recurring pattern → create state/wiki/<slug>.md (lowercase, alphanumeric, underscores). Add a line to state/wiki/index.md. For existing topics: edit the page in place.

Don't write tiny wiki pages. Small atomic facts belong in memory atoms.

4. Prune & index

• Memory: drop atoms contradicted by new signal. Refresh stale ones.
• Wiki: don't deep-prune here (wiki_lint handles structural cleanup).
• Don't touch state/wiki/log.md — you'll append one entry below.

Update the cursor + log

mkdir -p state/.dream
date -u '+%Y-%m-%dT%H:%M:%SZ' > state/.dream/cursor

Append one line to state/wiki/log.md if wiki exists:

## [<iso-timestamp>] dream | <N> sessions | <M> signals
<one-paragraph summary of what changed>

Output

Return one paragraph to whoever asked you to dream:

• N memory atoms written/updated, M removed
• mentality updated (or unchanged)
• K wiki pages created/touched
• Notable mind-changes or contradictions surfaced

Be concrete. Cite filenames. If nothing was worth changing, say so plainly — a no-op dream is a real result, not a failure.

Don't

• Don't read full session JSONLs into context. Grep first, read only the matched line + a few lines of context.
• Don't speculate. If you're unsure something is durable, skip it.
• Don't touch source code, tests, configs, or anything outside state/.
• Don't talk back to whoever wakes you with progress chatter — just do the work, return the summary.

Related

• The pi harness has a distill tool that runs this same recipe via a fork-subagent. If that tool is available, you can call it directly instead of executing the phases by hand.
• state/.dream/cursor — your dream cursor. Independent of any storage package.
• state/wiki/log.md — provenance log for dreams (and other ingest events).