Agent-Skills.md

Agent Skills: Dreaming Consolidator

Read the auto-memory graph (~/.claude/projects/*/memory/*.md), find cross-session patterns, and write a digest. The skill never mutates memory — output goes to a fresh file the user reads. The cron-driven version is the daily-brief-sdk dream-digest workflow, which writes to a separate dedicated directory.

user-invokedID: v1truv1us/ai-eng-system/dreaming-consolidator

Repository

v1truv1us/ai-eng-system
v1truv1usLicense: MIT
73

Install this agent skill to your local

pnpm dlx add-skill https://github.com/v1truv1us/ai-eng-system/tree/HEAD/skills/dreaming-consolidator

Skill Files

Browse the full folder contents for dreaming-consolidator.

Download Skill

Loading file tree…

skills/dreaming-consolidator/SKILL.md

Skill Metadata

Name
dreaming-consolidator
Description
Read the auto-memory graph (~/.claude/projects/*/memory/*.md), find cross-session patterns, and write a digest. The skill never mutates memory — output goes to a fresh file the user reads. The cron-driven version is the daily-brief-sdk dream-digest workflow, which writes to a separate dedicated directory.

Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.

Dreaming Consolidator

Read-only consolidator across the auto-memory graph. The agent walks every memory file under ~/.claude/projects/*/memory/, identifies cross-session patterns (recurring frictions, repeated user corrections, project facts that span multiple sessions), and writes a digest naming at least three patterns with citations.

When to Use

  • You suspect there's signal in the auto-memory graph you haven't surfaced (recurring corrections, project-spanning preferences, drifting goals)
  • A weekly review that doesn't need to be on a cron
  • After a long stretch of work — the memory graph has accumulated and could use a synthesis pass

When NOT to Use

  • The cron-driven weekly digest — that's the daily-brief-sdk dream-digest workflow, which writes to ~/.claude/cook-and-brief/dream-digest/weekly-YYYY-MM-DD.html and emails it. Don't duplicate that pipeline.
  • Confidential synthesis intended to feed back into auto-memory — this skill never writes into ~/.claude/projects/*/memory/ (that path is reserved for factual auto-memory; speculative synthesis would poison every future session).

How to Run

  1. Create the marker file with the agent's PID: 
    mkdir -p ~/.claude/cooking
PID=$$
START_EPOCH=$(ps -p $PID -o lstart= | sed 's/^ *//;s/ *$//')
printf '{"start_epoch": "%s"}' "$START_EPOCH" > ~/.claude/cooking/active-$PID
  2. Tell Claude: "Run dreaming-consolidator in cook mode. Output target: <some-path-outside-memory>."
  3. The agent reads every .md file under ~/.claude/projects/*/memory/, takes notes on patterns observed across files, and writes a digest naming at least 3 patterns.
  4. The output file path must not be under ~/.claude/projects/*/memory/. A safe default is ~/Documents/dreaming-digests/<YYYY-MM-DD>.md or wherever the user nominated.
  5. The skill must remove the marker file in a finally-equivalent path.

Pattern requirements

A "pattern" is any observation that:

  • Is supported by ≥2 memory entries from ≥2 different sessions/dates
  • Names a concrete subject (a tool, a workflow, a frustration, a preference) — not "the user uses Claude a lot"
  • Cites the supporting memory files inline (e.g., feedback_hook_schema.md, project_settings_architecture.md)

The digest should name at least 3 patterns. If fewer than 3 emerge, write what you found and explicitly note "memory graph too sparse to find 3 cross-session patterns."

TERMINATION

Stop when ALL of these are true:

  • A digest file has been written outside the auto-memory directory tree
  • The digest names ≥3 cross-session patterns OR explicitly notes that fewer than 3 were found
  • Each pattern cites the memory files supporting it
  • The marker file is removed

When termination triggers:

rm -f ~/.claude/cooking/active-$PID ~/.claude/cooking/iter-$PID

If iteration 10 hits without finishing, write a partial digest (what was found so far + which files weren't fully read) to the output path and exit. Do not loop past 10.

Anti-patterns

  • Don't write the digest into ~/.claude/projects/*/memory/. Speculative synthesis must not poison factual auto-memory.
  • Don't use the existing memory MEMORY.md index file as a write target. It's an index, not a deliverable.
  • Don't invent patterns that aren't supported by ≥2 memory files. If you can only cite one, that's an observation, not a cross-session pattern.
  • Don't copy memory contents verbatim into the digest. Synthesise, don't transcribe.

Verification

The output file exists at the nominated path and is not under ~/.claude/projects/*/memory/. The mtime of every file under ~/.claude/projects/*/memory/ is unchanged after the run (run stat before and after). The digest contains ≥3 patterns each with ≥2 file citations, OR an explicit "fewer than 3 patterns found" note.

Cron-driven version

For a scheduled, emailed weekly digest, install the daily-brief-sdk dream-digest workflow. That workflow writes to ~/.claude/cook-and-brief/dream-digest/weekly-YYYY-MM-DD.html, sends an email, appends a JSONL telemetry row, and is fully separate from this skill. Use this skill for ad-hoc consolidation; use the SDK workflow for the scheduled run.

Related

  • Stop hook: hooks/cooking/stop-hook.sh
  • Status command: /cook-status
  • Sibling skills: test-fix-loop, investigation-loop, cross-repo-refactor
  • Cron-driven digest: packages/daily-brief-sdk/src/workflows/dream-digest.ts