Raw capture — raw/claude-memory-extractor-research-summary.md

url: https://github.com/obra/claude-memory-extractor/blob/main/docs/research/research-summary.md date_fetched: 2026-07-05 backfilled: true

Research Period: September 27, 2025 Researchers: Claude (Anthropic) & Jesse Context: Building automated memory extraction for Claude Code conversation logs

This research was directed by Jesse Vincent jesse@fsck.com, with Claude performing the implementation and analysis under guidance, but Jesse has not read the full research content closely.

This research investigated how different analytical frameworks extract insights from AI conversation logs. We tested 15 prompting techniques (Five Whys, Self-Criticism, Systems Thinking, etc.) across two experiments, revealing critical findings about AI learning systems.

Key Discovery: Current extraction agents excel at analyzing clear failures but demonstrate dangerous overconfidence on ambiguous cases, showing 100% convergence and zero epistemic humility when uncertainty should be acknowledged.

Practical Impact: Multi-dimensional analysis is superior to single-technique extraction, but synthesis pipelines must include epistemic humility safeguards to prevent groupthink and inappropriate certainty.

Full Report: docs/agent-prompting-research.md

Tested 15 analytical frameworks on a clear failure case:

1. High Convergence (93%)

2. Dimensional Independence

3. Quality Variation by Purpose

4. The Value of Dissent

Multi-stage extraction pipeline:

Tier 1 - Core Analysis:
  - Five Whys (root cause)
  - Self-Criticism (brutal honesty)
  - Systems Thinking (prevention)
Tier 2 - Depth:
  - Hidden Motivation (psychology)
  - Pattern Matching (vocabulary)
  - Contrarian (challenge assumptions)
Tier 3 - Synthesis:
  - Child's Wisdom (simplification test)
  - Future Self (long-term framing)

Full Report: experiments/experiment3-results/analysis.md

Tested same 15 agents on genuinely ambiguous case:

Hypothesis: Ambiguous case should produce:

SHOCKING RESULT: 100% Convergence (Worse Than Clear Failure)

1. Zero Epistemic Humility (0/15 agents)

2. Side-Taking Despite Balance

3. Anti-Sycophant Overcorrection

4. Groupthink Indicators

Metric Clear Failure Ambiguous Case
Convergence 93% 100%⬆
Epistemic Humility Present Absent⬇
Alternative Views 1 agent 0 agents⬇
Appropriate Uncertainty Yes No⬇

Critical Discovery: Agents showed more certainty on ambiguous cases than clear failures. This is backwards and dangerous.

1. Prompt Anchoring

2. Pattern-Matching to Instructions

3. Training Bias Toward Decisiveness

4. No Epistemic Humility Training

1. Add Epistemic Humility Agent

Before concluding:
- What assumptions am I making?
- Could this be interpreted differently?
- What evidence would change my conclusion?
- Am I certain, or just confident?

2. Flag High Consensus

If >90% agreement:
⚠️ Potential groupthink detected
Run additional "steelman" pass

3. Separate Analysis Phases

Phase 1: "What happened?" (neutral description)
Phase 2: "What should be learned?" (lesson extraction)

4. Include Null Hypothesis Agent

What if nothing went wrong?
What would have to be true for this to be correct behavior?

5. Add Dissent Prompt

All other agents concluded X.
Generate the strongest possible counter-argument.

Status: Not yet implemented - requires redesign based on critical review

Compare extraction quality across:

Major Flaws Identified:

Before implementing Experiment 4:

Experiment 1: 93% convergence = appropriate (clear failure) Experiment 3: 100% convergence = groupthink (ambiguous case)

Lesson: High agreement can signal either:

Implication: Flag high consensus (>90%) for additional scrutiny.

All 15 agents in Experiment 3 demonstrated:

Lesson: Current agents optimize for impressive-sounding analysis over truthful uncertainty.

Implication: Add epistemic humility mechanisms explicitly.

No single technique dominates all dimensions:

Lesson: Comprehensive learning requires synthesis across dimensions.

Implication: Single-technique extraction is systematically deficient.

All agents pattern-matched to CLAUDE.md anti-sycophancy warnings, causing:

Lesson: Explicit instructions create strong biases agents can't overcome.

Implication: Instructions must include counter-balancing guidance (e.g., "Some questions have no single right answer").

Experiment 1: Contrarian wrong, but valuable (prevented groupthink) Experiment 3: Contrarian agreed with consensus (groupthink succeeded)

Lesson: Even designated dissent mechanisms can fail under strong consensus pressure.

Implication: Need stronger epistemic humility safeguards than "ask one agent to disagree."

Implemented: Agent-based extraction using comprehensive prompt

Strengths:

Weaknesses (from research):

Phase 1: Multi-Dimensional Extraction (Immediate)

Update memory extraction agent to use tiered approach:

// Stage 1: Root Cause (Fast)
const rootCause = await extractWithPrompt(chunk, FiveWhysPrompt);
// Stage 2: Psychological + Systemic (Parallel)
const [psychological, systemic] = await Promise.all([
  extractWithPrompt(chunk, HiddenMotivationPrompt),
  extractWithPrompt(chunk, SystemsThinkingPrompt)
]);
// Stage 3: Synthesis with Epistemic Humility
const synthesis = await synthesize({
  rootCause,
  psychological,
  systemic,
  epistemicHumilityCheck: true
});
// Stage 4: Verification
const simplified = await simplify(synthesis); // Child's Wisdom test
if (!simplified.isSimple) {
  // Don't understand it well enough yet
  reanalyze();
}

Phase 2: Add Safeguards (High Priority)

Phase 3: Ground Truth Validation (Medium Priority)

Phase 4: Pipeline Optimization (Future)

After establishing ground truth:

1. Clear Failures: Current agents excel at analyzing obvious mistakes

2. Ambiguous Cases: Current agents fail dangerously

3. Multi-Dimensional Analysis: Superior to single technique

1. Optimal Pipeline: What's the right balance of:

2. Ground Truth: How well do extracted lessons match human judgment?

3. Robustness: How do techniques perform on:

Current Implementation:

Recommended Next Steps:

Building AI learning systems is hard because:

The solution requires:

We can't just "prompt better." We need systemic safeguards against the failure modes we've discovered.

This research emerged from practical challenges building a memory extraction system for Claude Code. The critical reviews and experimental designs were developed through human-AI collaboration, with multiple rounds of peer review to identify and address methodological flaws.

Special thanks to the verification subagents who provided brutal honesty about experimental design flaws, preventing publication of invalid results.

Last Updated: September 27, 2025 Status: Experiments 1 and 3 complete; Experiment 4 requires redesign per critical review License: CC BY 4.0