Claude Cognitive

Working memory for Claude Code — persistent context and multi-instance coordination

The Problem

Claude Code is powerful but stateless. Every new instance:

• Rediscovers your codebase from scratch
• Hallucinates integrations that don't exist
• Repeats debugging you already tried
• Burns tokens re-reading unchanged files

With large codebases (50k+ lines), this becomes painful fast.

The Solution

Claude Cognitive gives Claude Code working memory through two complementary systems:

1. Context Router

Attention-based file injection with cognitive dynamics:

• HOT (>0.8): Full file injection - active development
• WARM (0.25-0.8): Headers only - background awareness
• COLD (<0.25): Evicted from context

Files decay when not mentioned, activate on keywords, and co-activate with related files.

2. Pool Coordinator

Multi-instance state sharing for long-running sessions:

• Automatic mode: Detects completions/blockers from conversation (every 5min)
• Manual mode: Explicit pool blocks for critical coordination
• Works with persistent sessions (days/weeks), not just short bursts

Results

Token Savings:

• Cold start: 79% (120K → 25K chars)
• Warm context: 70% (80K → 24K chars)
• Focused work: 75% (60K → 15K chars)

Average: 64-95% depending on codebase size and work pattern.

Developer Experience:

• ✅ New instances productive in first message
• ✅ Zero hallucinated imports/integrations
• ✅ No duplicate work across 8+ concurrent instances
• ✅ Persistent memory across days-long sessions

Validated on:

• 1+ million line production codebase (3,200+ Python modules)
• 4-node distributed architecture
• 8 concurrent Claude Code instances
• Multi-day persistent sessions

Quick Start

1. Install Scripts

# Clone to your home directory
cd ~
git clone https://github.com/GMaN1911/claude-cognitive.git .claude-cognitive

# Copy scripts
cp -r .claude-cognitive/scripts ~/.claude/scripts/

# Set up hooks (adds to existing config)
cat .claude-cognitive/hooks-config.json >> ~/.claude/settings.json

Note: The repo contains a .claude-dev/ directory for development/dogfooding purposes only. Do not copy this to your projects—it's not part of the user-facing installation. Use your own project-local .claude/ directory instead (see step 2).

2. Initialize Your Project

cd /path/to/your/project

# Create .claude directory
mkdir -p .claude/{systems,modules,integrations,pool}

# Copy templates
cp -r ~/.claude-cognitive/templates/* .claude/

# Edit .claude/CLAUDE.md with your project info
# Edit .claude/systems/*.md to describe your architecture

3. Set Instance ID

# Add to ~/.bashrc for persistence:
export CLAUDE_INSTANCE=A

# Or per-terminal:
export CLAUDE_INSTANCE=B

4. Verify It's Working

# Start Claude Code
claude

# First message - check for context injection:
# Should see: "ATTENTION STATE [Turn 1]" with HOT/WARM/COLD counts

# Query pool activity:
python3 ~/.claude/scripts/pool-query.py --since 1h

5. Create Keywords Config (Required)

Create .claude/keywords.json in your project root:

cp ~/.claude-cognitive/templates/keywords.json.example .claude/keywords.json

Edit to match your project's documentation files and relevant keywords.

Full setup guide: SETUP.md Customization guide: CUSTOMIZATION.md

Project Configuration

Create .claude/keywords.json in your project root to define project-specific keywords:

{
  "keywords": {
    "path/to/doc.md": ["keyword1", "keyword2", "phrase to match"]
  },
  "co_activation": {
    "path/to/doc.md": ["related/doc.md"]
  },
  "pinned": ["always/warm/file.md"]
}

Keywords: Map documentation files to trigger words. When any keyword appears in your prompt (case-insensitive), the file becomes HOT.

Co-activation: When a file activates, related files get a score boost.

Pinned: Files that should always be at least WARM.

The router checks for config in this order:

1. .claude/keywords.json (project-local)
2. ~/.claude/keywords.json (global fallback)
3. Empty defaults (no activation)

How It Works

Context Router

Attention Dynamics:

User mentions "orin" in message
    ↓
systems/orin.md → score = 1.0 (HOT)
    ↓
Co-activation:
  integrations/pipe-to-orin.md → +0.35 (WARM)
  modules/t3-telos.md → +0.35 (WARM)
    ↓
Next turn (no mention):
  systems/orin.md → 1.0 × 0.85 decay = 0.85 (still HOT)
    ↓
3 turns later (no mention):
  systems/orin.md → 0.85 × 0.85 × 0.85 = 0.61 (now WARM)

Injection:

• HOT files: Full content injected
• WARM files: First 25 lines (headers) injected
• COLD files: Not injected (evicted)

Pool Coordinator

Automatic Mode:

Instance A completes task
    ↓
Auto-detector finds: "Successfully deployed PPE to Orin"
    ↓
Writes pool entry:
  action: completed
  topic: PPE deployment to Orin
  affects: orin_sensory_cortex/
    ↓
Instance B starts session
    ↓
Pool loader shows:
  "[A] completed: PPE deployment to Orin"
    ↓
Instance B avoids duplicate work

Manual Mode:

```pool
INSTANCE: A
ACTION: completed
TOPIC: Fixed authentication bug
SUMMARY: Resolved race condition in token refresh. Added mutex.
AFFECTS: auth.py, session_handler.py
BLOCKS: Session management refactor can proceed
```

History Tracking (v1.1+)

Claude Cognitive now remembers its own attention. Every turn is logged with structured data showing which files were HOT/WARM/COLD and how they transitioned between tiers.

Why This Matters

The router always computed attention scores. Now they persist as queryable history:

• Replay development trajectories - "How did we stabilize the PPE last week?"
• Identify neglected modules - "Which files got ignored during the sprint?"
• Debug attention behavior - "Why didn't convergent.md activate when I mentioned convergence?"

View History

# Last 20 turns
python3 ~/.claude/scripts/history.py

# Last 2 hours
python3 ~/.claude/scripts/history.py --since 2h

# Filter by file pattern
python3 ~/.claude/scripts/history.py --file ppe

# Show only tier transitions
python3 ~/.claude/scripts/history.py --transitions

# Summary statistics
python3 ~/.claude/scripts/history.py --stats

# Filter by instance
python3 ~/.claude/scripts/history.py --instance A

Example Output

============================================================
  2025-12-31
============================================================

[18:43:21] Instance A | Turn 47
  Query: refactor ppe routing tier collapse
  🔥 HOT: ppe-anticipatory-coherence.md, t3-telos.md
  🌡️  WARM: orin.md, pipeline.md
  ⬆️  Promoted to HOT: ppe-anticipatory-coherence.md
  ⬇️  Decayed to COLD: img-to-asus.md

[19:22:35] Instance A | Turn 48
  Query: what divergence dynamics?
  🔥 HOT: divergent.md, t3-telos.md, cvmp-transformer.md
  🌡️  WARM: pipeline.md, orin.md (+3 more)
  ⬆️  Promoted to HOT: divergent.md

Statistics View

python3 ~/.claude/scripts/history.py --stats --since 7d

╔══════════════════════════════════════════════════════════════╗
║                    ATTENTION STATISTICS                      ║
╚══════════════════════════════════════════════════════════════╝

Total turns: 342
Time range: 2025-12-24 to 2025-12-31

Instances: {'A': 156, 'B': 98, 'default': 88}

Most frequently HOT:
   87 turns: pipeline.md
   65 turns: t3-telos.md
   43 turns: orin.md
   38 turns: ppe-anticipatory-coherence.md
   22 turns: divergent.md

Most promoted to HOT:
   23 times: ppe-anticipatory-coherence.md
   18 times: divergent.md
   12 times: convergent.md

Busiest days:
  2025-12-30: 156 turns
  2025-12-29: 98 turns
  2025-12-28: 88 turns

Average context size: 18,420 chars

History Entry Structure

Each turn logs:

{
  "turn": 47,
  "timestamp": "2025-12-31T18:43:21Z",
  "instance_id": "A",
  "prompt_keywords": ["refactor", "ppe", "routing", "tier"],
  "activated": ["ppe-anticipatory-coherence.md"],
  "hot": ["ppe-anticipatory-coherence.md", "t3-telos.md"],
  "warm": ["orin.md", "pipeline.md"],
  "cold_count": 12,
  "transitions": {
    "to_hot": ["ppe-anticipatory-coherence.md"],
    "to_warm": ["orin.md"],
    "to_cold": ["img-to-asus.md"]
  },
  "total_chars": 18420
}

File: ~/.claude/attention_history.jsonl (append-only, one entry per turn)

Retention: 30 days (configurable in context-router-v2.py)

Architecture

claude-cognitive/
├── scripts/
│   ├── context-router-v2.py      # Attention dynamics + history logging
│   ├── history.py                # History viewer CLI (v1.1+)
│   ├── pool-auto-update.py       # Continuous pool updates
│   ├── pool-loader.py            # SessionStart injection
│   ├── pool-extractor.py         # Stop hook extraction
│   └── pool-query.py             # CLI query tool
│
├── templates/
│   ├── CLAUDE.md                 # Project context template
│   ├── systems/                  # Hardware/deployment
│   ├── modules/                  # Core systems
│   └── integrations/             # Cross-system communication
│
└── examples/
    ├── small-project/            # Simple example
    ├── monorepo/                 # Complex structure
    └── mirrorbot-sanitized/      # Real-world 50k+ line example

Hooks:

• UserPromptSubmit: Context router + pool auto-update
• SessionStart: Pool loader
• Stop: Pool extractor (manual blocks)

State Files:

• .claude/attn_state.json - Context router scores
• .claude/pool/instance_state.jsonl - Pool entries

Strategy: Project-local first, ~/.claude/ fallback (monorepo-friendl