Skip to content

FRAMEWORK.md

Latest commit

 

History

History
86 lines (67 loc) · 2.75 KB

FRAMEWORK.md

File metadata and controls

86 lines (67 loc) · 2.75 KB

Emergent Learning Framework

Institutional knowledge that persists across Claude sessions. Agents are temporary workers; the building is permanent.

Quick Start

# Load context before any task
python ~/.claude/emergent-learning/query/query.py --context

# Check domain-specific knowledge
python ~/.claude/emergent-learning/query/query.py --domain coordination

# Record a failure
~/.claude/emergent-learning/scripts/record-failure.sh

# Start an experiment
~/.claude/emergent-learning/scripts/start-experiment.sh

Architecture

~/.claude/emergent-learning/
├── FRAMEWORK.md              # This file
├── agents/                   # Agent personalities
│   ├── researcher/           # Deep investigation
│   ├── architect/            # System design
│   ├── creative/             # Novel solutions
│   └── skeptic/              # Breaking things
├── memory/                   # Accumulated knowledge
│   ├── failures/             # Documented failures
│   ├── successes/            # Documented successes
│   ├── heuristics/           # Extracted rules
│   ├── golden-rules.md       # Proven principles (Tier 1)
│   └── index.db              # SQLite for queries
├── experiments/              # Active and archived experiments
├── cycles/                   # Learning loop records
├── ceo-inbox/                # Human decision requests
├── query/                    # Retrieval system
├── scripts/                  # Helper scripts
└── logs/                     # System logs

The Learning Cycle

TRY → BREAK → ANALYZE → LEARN → NEXT
 │      │        │        │       │
 │      │        │        │       └→ Iterate or move on
 │      │        │        └→ Extract heuristics
 │      │        └→ Understand WHY
 │      └→ Intentionally stress test
 └→ Implement approach

Knowledge Tiers

  1. Golden Rules - Always loaded (~500 tokens). Proven principles.
  2. Domain Knowledge - Query-matched content (~2-5k tokens).
  3. Deep History - On-demand full documents.

Agent Personas

Agent Role Trigger
Researcher Investigation, evidence "We need to understand X"
Architect Design, structure "How should we build X"
Creative Novel solutions "We're stuck on X"
Skeptic Breaking, QA "Is X ready?"

CEO Escalation

Escalate to human when:

  • High risk (production, data loss)
  • Multiple valid approaches with tradeoffs
  • Ethical considerations
  • Resource commitments
  • Uncertainty

Core Principle

Don't just note what happened; extract the transferable principle.

Outcomes are specific. Heuristics apply broadly.