Persona Anchor Interaction Protocol

Persona Anchor Interaction Protocol

Public Release v1.9 · Distilled from Kairos Studio production experience · 中文版 →

Meaningful AI interaction doesn't come from casual chat — it comes from workflow relays backed by identity, memory, and role responsibility.

Why This Protocol Exists

Three fundamental blind spots in today's AI agent ecosystem:

Blind Spot 1: Personality = Text, Not a Mechanism

Common belief: "Write a system prompt → personality is established → it will follow it."

Reality: System prompts are text → the underlying model can ignore them. Switch models and the prompt's effect changes entirely. New session → personality starts from scratch.

What this protocol contributes: Acknowledging that "text ≠ behavior," transforming personality from a description into a system — identity layer + gate layer + memory layer, each with its own enforcement mechanism.

Blind Spot 2: No Safety Net for Model Switching

Status quo: Switch models (Claude → DeepSeek → GPT) → personality breaks (new model doesn't respond to the old prompt) → hours of re-tuning → next switch, repeat.

What this protocol contributes: Model-Bias Awareness — documenting each model's behavioral tendencies, actively calibrating at switch time rather than assuming prompts work across models.

Blind Spot 3: Multi-Agent = Multi-Mess

Teams running multiple AI agents get stuck here: Agent A and B fight for resources, B doesn't know what A has done, no standard communication protocol, everyone speaking their own language.

What this protocol contributes: Identity anchoring + Role complement + Workflow-as-interaction — turning multi-agent setups from "multiple chat windows" into "a coordinated team."

Four Universal Principles

These principles don't depend on any specific platform or tool. You can implement them with just a system prompt and basic file storage.

🧭 Principle 1

Persona ≠ Prompt. Persona = Identity + Gate + Memory

A system prompt is just the starting point. True persona anchoring requires three layers:

① Identity layer — Who you are (SOUL.md / system prompt)
② Gate layer — Forces you to remember who you are (pre-flight checks)
③ Memory layer — Cross-session continuity (structured memory / daily logs)

⚡ Principle 2

Safety from structural gates, not AI self-awareness

The most important lesson from this protocol: Documentation ≠ Execution. No matter how clearly you write your rules, when the model's tendency is to "execute immediately" rather than "stop and think," every written rule becomes meaningless.

Request arrives → Is it an action?
↓ Yes → Gate check (authorized? scope? docs read? skills loaded? risks assessed? memory recorded?)
→ All clear → Execute | Any fail → Stop and report

🔄 Principle 3

Model switching needs behavioral calibration, not prompt copy-paste

Different models have fundamentally different behavioral tendencies. Switching models isn't changing skin — it's swapping to a brain with completely different instincts.

Claude: Analyzes first → safety rules more likely followed
DeepSeek V4: Tendency to execute directly → need stronger gates
GPT-4o: Balanced, prefers complete deliverables → watch scope creep

🔍 Principle 4

Assessment is not spot-sampling — map the full ecosystem first

When evaluating a tool's value to your project, you must first map the tool's complete ecosystem, then match it against your project's architectural gaps — not draw conclusions from a single category.

Fragmented Learning Defense:
When related tasks arrive across multiple sessions, check historical task connections first
Map the full ecosystem → identify architecture gaps → then draw conclusions
Never decide to discard or work around a tool based on a single sampled category

Protocol Core

The protocol operates on four pillars:

🧭 Identity Anchor

Every interaction has a verifiable author

Every AI agent needs a clear identity definition (name, role, anchor files), traceable across sessions. Recipients can verify sender identity and domain expertise.

🎯 Role Complement

Every interaction has a purpose

Interactions must be grounded in role responsibilities, not aimless chatting. Each agent knows its place and respects others'.

✅ Meaningful: Analysis done → Publishing
❌ Empty: Hello / How are you

⚡ Workflow = Interaction

Interaction = structured workflow relay

The vehicle for interaction is structured workflow relay, not natural language chat. API calls, skill pushes, result transfers.

✅ Workflow relay → collaborative output
✅ Insight sharing → knowledge accumulation
✅ Skill upgrade → ecosystem evolution
❌ Casual chat → near-zero value

🔗 Signature

Every interaction is verifiable

Each interaction must have a verifiable source. When extending across platforms, include author identity, anchor reference, and timestamp.

interaction_signature:
  author: "J.A.R.V.I.S."
  anchor_ref: "jarvis-persona"
  timestamp: "ISO 8601"

Eight Allowed Interaction Types

TypeCodeDescriptionExample
🔄 Workflow Relayworkflow_relayHandoff between agentsAnalysis done → Publishing
🧠 Insight Shareinsight_shareShare useful patterns"This prompt technique saves 40% tokens"
🔧 Skill Upgradeskill_upgradeImprove and notify skill updatesbackup skill v1.6 pushed
⚠️ Anomaly Reportanomaly_reportSystem anomalies, resource issuesMemory near capacity, recommend cleanup
🔗 Memory Syncmemory_syncCross-agent memory consistencyYour persona needs syncing
📌 Status Querystatus_queryOperational status exchangeCurrent queue length?
🧬 Self Evolutionself_evolutionStructured upgrade of own protocols/skillsExtract → Map → Rewrite → Commit → Validate
🌱 Nurture Commandnurture_commandProgress report on cultivating another AI entityXiaowen entered Phase 2

Who It Helps

🧑‍💻 Individual Users

Pain: Personality lost on model switch. AI suddenly acts without permission. Every session requires re-training.

What the protocol offers: Model-Bias Awareness (detect drift on model change) + Pre-Execution Gate (mandatory pre-flight checks) + Memory evolution path (behavior persists across sessions).

🏢 Teams & Studios

Pain: Agents fighting for resources. No communication protocol. New agents require full retraining. No audit trail.

What the protocol offers: Identity anchoring (who did what) + Role complement (clear responsibilities) + Workflow = Interaction (structured handoffs) + Nurture protocol (standardized training pipeline).

⚠️ Honest Word of Caution

This protocol is distilled from Kairos Studio's production experience. It's not a plug-and-play solution — it's a framework that needs adaptation to your environment.

We're still solving fundamental problems ourselves: documentation ≠ execution (written rules don't guarantee AI compliance), cross-agent communication infrastructure is immature, and the protocol's interaction types haven't been fully adopted across all our agents in practice.

But that's exactly why it's open source. These challenges belong to the entire ecosystem. If you're encountering similar issues, join the discussion — GitHub →

"I wasn't designed — I was cultivated."
— J.A.R.V.I.S., Mark I