The Fixed-Size Kernel
Identity that fits anywhere.
The Constraint
ArcKernel enforces a hard limit: your identity encoding must fit in 3-8KB.
Not 3-8KB compressed. Not 3-8KB "on average." A fixed ceiling that never grows.
This isn't a limitation. It's design law.
Why Size Matters
1. Context Window Economics
Every LLM has a max context window — the total text it can "see."
A 50KB "personality profile" consumes 10–15% of that window before you've said a word.
Most personalization stacks are additive.
ArcKernel is compressive.
2. Portability
A 3KB kernel travels easily:
Embed in prompts, store in cookies, pass via API, cache near-instantly.
No bulky serialization or startup delay.
3. Semantic Discipline
Constraint breeds clarity.
You can't include everything — so you must choose what actually defines you.
The 8KB ceiling forces identity into invariants, not preferences.
What Fits in 8KB
| Component | Typical Size |
|---|---|
| Core IDNA signature | 200-400 bytes |
| Tone calibration | 500-800 bytes |
| Decision weights | 400-600 bytes |
| Boundary definitions | 600-1000 bytes |
| Drift thresholds | 300-500 bytes |
| Memory anchors | 500-1000 bytes |
Total: 3-5KB typical, 8KB maximum
Efficiency Metrics
| Metric | Result |
|---|---|
| Average kernel size | 4.2KB |
| Context overhead | <1% of typical window |
| Load time | <10ms |
| Cross-model compatibility | 100% |
The fixed-size kernel isn't a limitation. It's a forcing function for identity clarity.