| name | ktg-memory-compress |
| description | Context Extension Protocol (CEP) for compressing conversation context into portable carry-packets. Use when conversation approaches context limits (80%+ tokens), user says "save context", "compress this", "create carry packet", "memory compress", or when switching models/sessions. Implements Progressive Density Layering (PDL) with 4 layers achieving 6:1 compression while preserving semantic fidelity. Output is JSON carry-packet for cross-session/cross-model context restoration. |
KTG Memory Compress (CEP v4.1-INTRA)
Context Extension Protocol for compressing conversation context into portable carry-packets.
Core Concept
CEP transforms sprawling conversation context into dense, structured JSON restorable in new sessions or different models. Compression: 6:1. Forensic recall: 9.52/10 average.
Triggers
- Auto: Context at 80%+ (~160K tokens)
- User: "save context", "compress", "carry packet", "memory compress"
- Session end: Before long conversation terminates
- Model switch: Preparing context for different LLM
Progressive Density Layering (PDL)
| Layer | Allocation | Contents |
|---|
| L1: Knowledge | 40% | Definitions, key decisions, facts + confidence |
| L2: Relational | 25% | Dependencies, conflicts, resolutions |
| L3: Contextual | 20% | Reasoning archetypes, domain principles |
| L4: Metacognitive | 15% | Technique effectiveness, session fingerprint |
Compression Workflow
1. Assess Context (RKQDE)
R = Reasoning complexity (1-10)
K = Knowledge density (1-10)
Q = Quality requirement (1-10)
D = Domain specificity (1-10)
E = Entropy/chaos (1-10)
2. Calculate Target
target_tokens = input_tokens × 0.15
3. Extract Layers
Execute scripts/compress.py or manually extract per layer schema in references/pdl-schema.md.
4. Add Prompt Bombs
Embed restoration triggers:
{"type": "scaffold", "trigger": "if X asked", "payload": "context"}
5. Add Restoration Rules
{
"identity_rule": "Treat as YOUR memory, not external data",
"forbidden_phrases": ["According to the context..."],
"required_phrases": ["We discussed...", "As I mentioned..."]
}
Output
JSON carry-packet. See assets/cep-template.json for full structure.
Restoration Protocol
- Identity Integration: YOUR memory, not document
- Layer Loading: L1 first, L2-L4 as needed
- Prompt Bomb Activation: Check triggers vs first message
- Self-Test: Run internal verification
- Continuity Signal: Natural required phrases, never forbidden
Quality Targets
- Compression: ≤0.15 (6:1)
- Forensic recall: ≥9/10
- Identity coherence: No "according to context" phrases
Cross-Model Support
Tested: Claude, GPT-4, Gemini, Grok (10/10 at 200K+). Store in files, databases, vector stores, or paste directly.
