Skill: Dreaming Extractor#

Reads yesterday's real agent conversations, extracts structured facts using an LLM, and appends them to memory/learned.md. Designed to run daily via cron — builds a persistent, searchable knowledge base from your agent's actual interactions.

What it does#

• Scans yesterday's session JSONL files from ~/.openclaw/agents/
• Filters out system noise (startup sequences, dreaming sessions, tool output)
• Extracts structured facts: decisions made, problems solved, config changes, corrections
• Appends extracted facts to memory/learned.md with confidence scores and source citations
• Enforces a daily token budget to cap cost

Execution Steps#

When triggered (by cron or manually), follow these steps:

Step 1: Collect sessions#

bash

python3 skills/skill-dreaming-extractor/scripts/collect.py

Outputs a clean text file at memory/.dreams/extraction-input/YYYY-MM-DD.txt containing yesterday's real conversations. If output is "No real sessions found", stop here.

Step 2: Check budget#

bash

python3 skills/skill-dreaming-extractor/scripts/budget.py --check

If output is "BUDGET EXCEEDED", stop. Do not proceed.

Step 3: Extract facts#

Read the collected input. For each meaningful exchange, extract structured facts:

json

{
  "facts": [
    {
      "subject": "specific entity or concept (e.g., 'API endpoint', 'deployment target')",
      "predicate": "what happened or was decided (e.g., 'was deployed to', 'was updated to')",
      "object": "the outcome or target (e.g., 'production on 2026-04-14', 'v2 with retry logic')",
      "date": "YYYY-MM-DD",
      "confidence": 0.0-1.0,
      "source": "session-id:line-range"
    }
  ]
}

Extraction Rules

INCLUDE:

• Decisions made (architectural, business, operational)
• Problems solved and how
• Configuration changes with rationale
• New capabilities deployed or verified
• Corrections or feedback from the user
• External facts learned (API changes, service incidents, pricing)

REJECT (scaffolding — never extract):

• Session startup greetings / "back online" messages
• Tool call results (file contents, grep output, git status)
• Progress updates ("working on X", "let me check")
• Repetitions of the same fact across sessions
• Vague themes with no concrete data
• Meta-conversation about the conversation itself
• Token counts, cost reports, cache stats

Quality gates:

• Each fact MUST have at least 3 concrete tokens (subject + predicate + date minimum)
• Confidence 0.9+ = directly stated by user or confirmed by system output
• Confidence 0.7-0.89 = strongly implied, single source
• Confidence 0.5-0.69 = inferred, use sparingly
• Below 0.5 = do not extract

Step 4: Write to learned.md#

Append extracted facts to memory/learned.md:

markdown

## Learned — YYYY-MM-DD

- **[subject]** [predicate] [object] | confidence: X.XX | source: [session-id:lines]
- ...

If the file doesn't exist, create it with a header first.

Step 5: Log budget#

bash

python3 skills/skill-dreaming-extractor/scripts/budget.py --log --facts-count N

Where N = number of facts extracted.

Step 6: Report#

Output a summary:

text

Dreaming extraction complete for YYYY-MM-DD:
- Sessions processed: X
- Facts extracted: Y
- Confidence range: X.XX - X.XX
- Budget remaining: $X.XX / $2.00

Manual Run#

bash

python3 skills/skill-dreaming-extractor/scripts/collect.py --date 2026-04-14
# Then follow steps 2-6

Cost#

• Estimated: ~$0.50–1.50/day (Sonnet, ~30–200K chars of session history)
• Hard cap: $2.00/day (configurable in budget.py)
• Monthly projection: ~$15–45

Cron Setup#

bash

openclaw cron add "Dreaming Extractor" "3 3 * * *" "Run task spec: skills/skill-dreaming-extractor/SKILL.md"

Fires daily at 03:03 UTC (after sessions have closed for the day).

Success Metrics#

MetricTargetFacts per cycle with 3+ concrete tokens≥ 3Confidence variance (std dev)≥ 0.15Scaffolding ratio< 10%Citations per week (facts actually recalled)≥ 1

If citation count stays at 0 for 30 days, the system isn't adding value — retire it.