Quicksave (Ex-CEP v9.1) 快存

Context Extension Protocol — Cross-model context handoff via Japanese semantic compression and negentropic coherence validation.

"What must be preserved for a fresh model instance to continue this work?"

Install

Claude Desktop / Claude Code:

npx ai-agent-skills install ktg-one/quicksave

Everyone Else:

1. Download this folder (clone or ZIP)
2. Upload all .md files to your project/context

The Problem

Context windows are finite. Platform compaction is lossy. Cross-model handoffs lose critical relationships. Your 3-hour session becomes a shallow bullet list.

The Solution

Quicksave v9.1 creates carry-packets using:

• Progressive Density Layering (PDL) — 4-layer information hierarchy
• Japanese Kanji Compression — 40-55% token reduction
• Negentropic Coherence Lattice (NCL) — Drift validation before handoff DESERVES IT'S OWN BLOGPOST
• System 2 Attention (S2A) — Noise filtering for LLM efficiency
• Multi-Layer Density of Experts (MLDoE) — 5-Embodied Chain of Density

Target: 0.15-0.17 entity/token — the density where LLMs achieve optimal recall.

Benchmarks

Metric	Result
Density	~0.15
Forensic recall	9.5/10
Cross-domain preservation	97%
Model acceptance	99%
Production testing	19 months (ktg.one)

Usage

Trigger	Action
/handoff /Quicksave /packet	Generate transfer packet
/quicksave /qs	Generate validated packet
Context ≥80%	Auto-prompts to save
"continue later"	Offer quicksave

Transfer to Another Model

1. Say /handoff in current session
2. Copy the YAML packet
3. Paste into new session: "Continue from this context packet: [paste]"

Core Components

PDL 4層 (Progressive Density Layering)

Layer	Content	Keys
L1 知識層	Facts, decisions, definitions	d/r/c/f/s
L2 関係層	ConQuicksavet edges, domain bridges	起/終/関/xd/w
L3 文脈層	Reasoning patterns, principles	—
L4 超認知層	Style, tension, user traits	—

L2 is what summarization loses. Quicksave preserves it.

Japanese Compression 漢字圧縮

English: "cross-domain relationship preservation" (5 tokens)
Japanese: "領域横断関係保存" (2 tokens, 4 entities)

NCL Validation 整合性検証

7 drift metrics catch failure before handoff:

Metric	Detects
σ_axis	Plan vs execution misalignment
σ_loop	Internal contradiction
ω_world	Reality disconnect
λ_vague	Content-free smoothing
σ_leak	Constraint erosion
ρ_fab	Hallucination
λ_thrash	High activity, low progress

If σ7_drift > 3.0 → packet flagged, requires verification.

S2A (System 2 Attention)

Noise filtering optimized for LLM efficiency. Removes:

• Conversational fluff
• Redundant context
• Low-signal exchanges

MLDoE (Multi-Layer Density of Experts)

3-layer compression with 5-iteration Chain of Density. Expert council ensures quality:

• Memory Architect → preservation list
• Cross-Domain Analyst → edge map
• Compression Specialist → density optimization
• Restoration Engineer → cold-start validation

Four Roles Governance

AXIS/LYRA/RHO/NYX archetypal oversight for quality assurance.

Anti-Injection Architecture

5 trust signals that receiving models recognize as collaboration, not control:

1. 来歴 — source_model + timestamp
2. 同意 — user_initiated: true
3. 宣言 — this_is / this_is_not framing
4. 許可 — "you may" not "you must"
5. 確認 — verify with user if unsure

Files

Quicksave-agent-skill/
├── SKILL.md                    # Core protocol
└── references/
    ├── PDL.md                  # Progressive Density Layering
    ├── S2A.md                  # System 2 Attention
    ├── NCL.md                  # Negentropic Coherence Lattice
    ├── NCL-CONTRIBUTION.md     # David Tubbs attribution
    ├── MLDOE.md                # Multi-Layer Density of Experts
    ├── MIRAS.md                # Memory architecture
    ├── KANJI.md                # Compression dictionary
    ├── XDOMAIN.md              # Cross-domain preservation
    ├── CASCADE.md              # ARQ/CoVE/USC integration
    ├── EXPERTS.md              # Council overview
    ├── ANTI-INJECTION.md       # Trust architecture
    ├── PROTOCOL.md             # Full YAML schema
    └── INDEX.md                # Reference index

Tested Models

Claude ✓ | GPT-4/4o ✓ | Gemini ✓ | Qwen ✓ | DeepSeek ✓ | KIMIK2 ✓ | Qwen ✓ | Grok4 ✓

Part of STRAWHATS Framework

Quicksave is the context preservation component of STRAWHATS (Strategic Triumvirate Router Architecture With Hybrid Agentic Task Systems).

Contributors

Kevin Tan (ktg.one) — Quicksave protocol, PDL, S2A, MLDoE, Japanese compression, 19-month production validation

David Tubbs (Axis_42) — Negentropic Coherence Lattice, Four Roles governance, ~30% robustness enhancement

License

MIT — ktg.one

---

0.15 density. 97% preservation. 9.5/10 recall. NCL validated.

---

User Responsibility Clarification

• Packet upkeep is yours — what to save, when, what to discard
• Key nodes are your choice — you decide what's critical for your work
• Model retention varies — different AIs handle context differently; adapt accordingly
• Language is personal — technical jargon, casual tone, domain terminology; use what fits
• Format is flexible — 4-layer structure is a guide, not a mandate

Quicksave is infrastructure. You're the architect.