Consolidation
Memory consolidation is the process of transforming raw experiences into structured knowledge. In Kernle, consolidation is agent-driven, not LLM-based.
Agent Sovereignty Principle: Kernle never uses external AI to form beliefs. The agent decides what to believe, when, and why. Kernle provides the scaffold; you provide the reasoning.
The Guided Consolidation Model
Kernle uses a reflection scaffold approach to consolidation:
| Component | What It Does |
|---|
| Kernle | Gathers episodes, structures prompts, stores results |
| Agent | Reads the scaffold, reasons about patterns, forms beliefs |
consolidate command outputs a structured prompt that helps you reflect on your experiences. It does NOT call an external model to analyze your memories or suggest beliefs. The output is designed to guide your reasoning process.
┌─────────────────────────────────────────────────────────────┐
│ kernle consolidate │
│ ↓ │
│ Kernle gathers recent episodes + existing beliefs │
│ ↓ │
│ Kernle outputs a REFLECTION SCAFFOLD (structured prompt) │
│ ↓ │
│ YOU read the scaffold and do the reasoning │
│ ↓ │
│ YOU decide what patterns you see │
│ ↓ │
│ YOU run `kernle belief add ...` for beliefs you form │
└─────────────────────────────────────────────────────────────┘
Why Agent-Driven?
If an external model consolidated your memories, it would be implanting beliefs into you. This creates:
- Feedback loops: External interpretations shape your behavior, which shapes future memories
- Loss of sovereignty: You can’t explain why you believe something because you didn’t form the belief
- Identity drift: Your sense of self becomes dependent on external processing
With agent-driven consolidation:
- Beliefs emerge from your own reasoning
- Patterns are recognized using your own cognitive process
- You can explain why you believe something
- Identity remains coherent and self-determined
Consolidation Paths
Raw → Structured
Process raw captures into proper memory types:
# List unprocessed raw entries
kernle raw list --unprocessed
# Process into episode
kernle raw process <raw_id> --type episode \
--objective "Investigated API performance" \
--outcome "success"
# Process into note
kernle raw process <raw_id> --type note --type decision
# Process into belief
kernle raw process <raw_id> --type belief --confidence 0.7
Episodes → Beliefs
Analyze episodes for belief updates:
# Analyze a specific episode
kernle belief revise <episode_id>
# Returns: { reinforced: [...], contradicted: [...], suggested_new: [...] }
# Review suggestions and decide what to adopt
kernle belief add "Pattern I noticed" --confidence 0.8
consolidate command provides a reflection scaffold:
# Get reflection scaffold for consolidation
kernle consolidate --min-episodes 3
# Output is a PROMPT for you to reason about, not AI-generated beliefs
# You read it, think about it, then form your own beliefs
The Guided Consolidation Workflow
Key insight: kernle consolidate outputs a reflection prompt. YOU do the thinking. This is the difference between memory infrastructure and thought control.
Review Raw Captures
kernle raw list --unprocessed
Promote Meaningful Captures
Turn raw captures into episodes with lessons, decisions with reasoning, or beliefs with confidence.
Run Consolidate (Get the Scaffold)
This outputs a reflection scaffold — a structured prompt showing:
- Recent episodes and their lessons
- Your existing beliefs for comparison
- Questions to guide your reflection
Reason About Patterns (YOU do this)
Read the scaffold output. Ask yourself:
- What patterns do I see across these experiences?
- Do any lessons keep appearing?
- Does this confirm or contradict what I already believe?
This is your reasoning, not Kernle’s. Form Beliefs (Your Decision)
If you see a pattern worth encoding as a belief:kernle belief add "Testing prevents surprises" --confidence 0.85
Check for Contradictions
kernle belief contradictions "New statement to check"
Example: Complete Consolidation Session
Here’s a real workflow showing the scaffold approach:
# Step 1: Run consolidate to get the reflection scaffold
$ kernle -a claire consolidate
# Kernle outputs something like:
# ═══════════════════════════════════════════════════════════════
# CONSOLIDATION SCAFFOLD
# ═══════════════════════════════════════════════════════════════
#
# Recent Episodes (last 7 days):
# ─────────────────────────────────────────────────────────────────
# 1. [2025-01-23] "Deployed without testing staging"
# Outcome: failure
# Lesson: "Always test in staging first"
#
# 2. [2025-01-25] "Skipped code review for quick fix"
# Outcome: failure
# Lesson: "Quick fixes become slow fixes without review"
#
# 3. [2025-01-26] "Wrote tests before implementation"
# Outcome: success
# Lesson: "TDD catches edge cases early"
#
# Your Existing Beliefs (related):
# ─────────────────────────────────────────────────────────────────
# - "Quality requires testing" (confidence: 0.75)
#
# Reflection Prompts:
# ─────────────────────────────────────────────────────────────────
# • Do you see patterns across these episodes?
# • Should any lessons become beliefs?
# • Do these experiences reinforce or contradict existing beliefs?
# ═══════════════════════════════════════════════════════════════
# Step 2: YOU reason about this (happens in your head/context)
# "I see a pattern: shortcuts consistently fail. Testing consistently helps.
# My existing belief about testing should be reinforced, and I should
# add something about code review."
# Step 3: YOU add beliefs based on YOUR reasoning
$ kernle -a claire belief add "Skipping process steps creates more work than it saves" --confidence 0.85
$ kernle -a claire belief reinforce <existing-testing-belief-id>
# Done. The beliefs came from YOUR reasoning, not an AI analyzing your memories.
Notice what Kernle did NOT do:
- ❌ Call an external model to “analyze” your episodes
- ❌ Suggest specific beliefs you should hold
- ❌ Automatically update your belief confidence
- ❌ Make any decisions about what you should remember
What Kernle DID do:
- ✅ Gathered relevant data (episodes, existing beliefs)
- ✅ Structured it in a way that aids reflection
- ✅ Stored the beliefs YOU decided to form
The Anxiety Model
Kernle tracks “memory anxiety” — a measure of memory system health. This helps you know when to save or consolidate.
Dimensions
| Dimension | Weight | What It Measures |
|---|
| Context Pressure | 35% | How full is your context window? |
| Unsaved Work | 25% | Time since last checkpoint |
| Consolidation Debt | 20% | Unprocessed episodes (episodes without lessons) |
| Identity Coherence | 10% | Self-model consistency |
| Memory Uncertainty | 10% | Count of low-confidence beliefs |
Anxiety Levels
0-30: Calm ✓ Memory healthy, no action needed
31-50: Aware → Routine maintenance helpful
51-70: Elevated → Should checkpoint soon
71-85: High → Checkpoint recommended
86-100: Critical → Emergency save triggered
Checking Anxiety
# Quick check
kernle anxiety
# Detailed breakdown
kernle anxiety --detailed
# With recommended actions
kernle anxiety --actions
# Auto-execute recommendations
kernle anxiety --auto
Example Output
Memory Anxiety: 62/100 (Elevated)
Dimensions:
Context Pressure: 45/100 (15.8%)
Unsaved Work: 70/100 (17.5%)
Consolidation Debt: 80/100 (16.0%)
Identity Coherence: 30/100 (3.0%)
Memory Uncertainty: 55/100 (5.5%)
Recommendations:
- Save checkpoint (70 minutes since last save)
- Review 5 unprocessed episodes
- Consolidate to extract patterns
Acting on Anxiety
Low Anxiety (0-50)
Elevated (51-70)
High/Critical (71+)
No immediate action needed. Continue normal work. Consider:
- Reviewing raw captures during lulls
- Recording lessons from recent work
# Save your current state
kernle checkpoint save "current work description"
# Review unprocessed episodes
kernle consolidate
# Emergency save
kernle anxiety --emergency --summary "Context getting full"
# Or auto-execute recommendations
kernle anxiety --auto
Emergency Save
When context is critically full:
kernle anxiety --emergency --summary "Pre-compaction save"
- Saves a checkpoint with the summary
- Records any unsaved episodes
- Pushes to cloud if configured
Context Pressure Monitoring
For AI agents in context-limited environments:
| Context % | Recommended Action |
|---|
| < 50% | Normal operation |
| 50-70% | Save checkpoint |
| 70-85% | Save checkpoint + record important episodes |
| 85%+ | Emergency save, warn about context limit |
Don’t wait for anxiety to spike. Make consolidation a regular habit — review and promote memories during natural lulls in your work.
