Feature: Dreaming + AutoResearch — autonomous knowledge consolidation with verification

[huichalaf]

Summary

Combine ClawCode's dreaming system with Karpathy's autoresearch pattern to create a self-improving knowledge agent that autonomously researches gaps detected during the day, validates findings, and only promotes verified knowledge to long-term memory.

Current dreaming promotes memories based on recall frequency. This proposal adds an active research loop during the REM phase that fills knowledge gaps and verifies information before promotion — shifting from "remember what was searched often" to "learn what was missing and verify it's correct."

Motivation

Today's dreaming system is passive — it only consolidates what the agent already encountered. But during daily use, the agent frequently hits knowledge gaps (low-score searches, unanswered questions, partial results). These gaps are detected but never acted upon.

Karpathy's autoresearch showed that a simple propose→test→keep/discard loop can run hundreds of validated improvements overnight with zero human intervention. The same pattern applies to knowledge consolidation:

	AutoResearch (Karpathy)	Dreaming + AutoResearch
Optimizes	val_bpb (training metric)	Knowledge coverage & accuracy
Modifies	train.py (code)	MEMORY.md (knowledge)
Validation	Did loss improve?	Is it verifiable and correct?
Keep/Discard	git commit / git revert	promote / discard snippet
Loop trigger	Continuous (while true)	Nightly (3 AM, integrated with dreaming)

Proposed Design

Gap Detection (during day)

Extend trackRecall to also track gaps — searches that returned zero or low-quality results:

interface KnowledgeGap {
  query: string;
  timestamp: string;
  resultCount: number;
  maxScore: number;        // 0 if no results
  gapType: "no_results" | "low_confidence" | "partial_match";
}

Store in memory/.dreams/knowledge-gaps.json, alongside existing short-term-recall.json.

Enhanced REM Phase — Research Loop

During the REM phase, for each detected gap:

for each gap in knowledge-gaps.json (sorted by frequency):
  1. RESEARCH — gather information
     ├─ WebSearch (if available)
     ├─ Codebase search (Grep/Read relevant files)
     ├─ Documentation lookup
     └─ Cross-reference with existing memory
  
  2. VALIDATE — verify findings
     ├─ Multi-source confirmation
     ├─ Code cross-reference (if domain is code-related)
     └─ Assign confidence_score (0.0 - 1.0)
  
  3. EVALUATE — keep or discard
     ├─ confidence >= 0.7 → KEEP → stage as Deep phase candidate
     └─ confidence < 0.7  → DISCARD → log reason in DREAMS.md

Research Candidate Schema

interface ResearchCandidate {
  query: string;                // the original gap
  sources: string[];            // where info was found
  snippet: string;              // synthesized knowledge
  confidenceScore: number;      // 0-1
  validationMethod: 
    | "code-crossref"           // confirmed against codebase
    | "web-confirm"             // confirmed via web search
    | "multi-source"            // multiple sources agree
    | "single-source";          // only one source (lower trust)
  kept: boolean;
  discardReason?: string;       // why it was discarded
}

Deep Phase Integration

Research candidates that pass validation feed into the existing Deep phase scoring alongside recall-based candidates. They get an additional signal:

Signal	Weight	Description
Research confidence	bonus	Verified knowledge gets a scoring boost

DREAMS.md Output

## REM Research — 2026-04-13 03:00

### Investigated 3 knowledge gaps

1. ✅ "navitaire change flight endpoint" (confidence: 0.89)
   - Sources: codebase (bookingModification.service.ts), web docs
   - Learned: POST /api/nsk/v4/booking/flights with {journeyKey, fareKey}
   - Validation: code cross-reference confirmed

2. ✅ "409 error re-quote" (confidence: 0.92)  
   - Sources: codebase (navitaire.service.ts), error logs
   - Learned: 409 occurs when PNR has active lock
   - Validation: code pattern confirmed

3. ❌ "ancillary pricing connections" (confidence: 0.45) — DISCARDED
   - Reason: insufficient sources, web results contradictory
   - Action: needs user clarification

Implementation Considerations

What stays deterministic

• Gap detection (trackRecall extension)
• Confidence scoring formula
• Keep/discard threshold
• Deep phase integration

What requires LLM

• Research synthesis (generating snippets from sources)
• Validation reasoning (why something is/isn't trustworthy)
• Dream diary narrative

Configuration

{
  "dreaming": {
    "autoresearch": {
      "enabled": false,
      "maxGapsPerNight": 5,
      "confidenceThreshold": 0.7,
      "sources": ["codebase", "memory", "web"],
      "maxResearchTimeMinutes": 10
    }
  }
}

Safety constraints

• Read-only by default — research never modifies code or external systems
• Confidence gating — only verified knowledge gets promoted
• Rate limiting — cap on API calls per dream cycle
• Transparency — everything logged in DREAMS.md for human review
• Opt-in — disabled by default, like dreaming itself

Compound Effect

Day 1: Agent learns about endpoint X. Day 5: when researching related topic Y, it already has context from day 1. Knowledge compounds like interest — each night's research builds on previous nights.

Over time, the agent develops domain expertise autonomously, filling gaps the user never explicitly taught it.

References

• karpathy/autoresearch — the original propose→test→keep/discard loop
• Karpathy on the future of autoresearch — "emulate a research community"
• ClawCode's existing dreaming system (lib/dreaming.ts) — the foundation this builds on
• OpenClaw's dreaming docs (docs/concepts/dreaming.md) — the upstream reference

Open Questions

1. Should research candidates go through the same 6-signal scoring as recall candidates, or have their own scoring pipeline?
2. Should the agent be allowed to use subagents for parallel research (like OpenClaw's narrative generation)?
3. How to handle research that contradicts existing MEMORY.md entries? (knowledge correction)
4. Should there be a "research backlog" for gaps that couldn't be resolved, similar to IMPORT_BACKLOG.md?

---

Happy to contribute an implementation PR if there's interest. This feels like a natural evolution of the dreaming system.