Skip to content

Latest commit

 

History

9 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BTP — Behavioral Transfer Protocol

Your AI gets better every conversation. No training. No fine-tuning. Just text files.

The problem

You correct your AI assistant ten times. It keeps making the same mistakes. New session — amnesia. Switch models — start over. Your corrections vanish.

What BTP does

Every correction you give becomes a permanent behavioral change — across sessions, across models, instantly.

You: "Don't give me five options. Just pick the best one."
     ↓
BTP detects the correction
     ↓
Compiles it into a format any model understands
     ↓
Next response: the model picks one and explains why
     ↓
Every future response: same improvement. Forever.

No retraining. No GPU. No API changes. Works on ChatGPT, Claude, Gemini, Llama, or any model that reads text.

How it works

Three layers, each teaching the model differently:

1. Rules — Hard boundaries

Never hallucinate. Say "I don't know" instead.

The model follows this literally. Binary compliance.

2. Pattern Interrupts — Override instincts

YOUR INSTINCT: Give 5 options and ask the user to pick.
WHAT ACTUALLY WORKS: Pick the best one. Explain why. Offer alternatives only if asked.
TRIGGER: Any time you're about to list more than 2 options.

Predicts what the model will naturally do and redirects it. Works on first read.

3. Correction Transcripts — Simulated experience

SITUATION: User asks for a database recommendation.
WRONG: "Here are some options: PostgreSQL, MySQL, MongoDB, SQLite, DynamoDB..."
CORRECTION: "Don't give me a menu. Pick one."
RIGHT: "PostgreSQL. Handles relational data well, strong ecosystem, scales to most workloads."

Shows the pattern of being corrected. The model recognizes it before making the same mistake.

Results

Tested: transferring 46 sessions of behavioral corrections from Claude Opus to GPT-5.1-codex-max.

Rules only:                ██░░░░░░░░  20%  — read the rules, ignored them
+ Pattern Interrupts:      ██████░░░░  67%  — acted on data instead of reporting
+ Correction Transcripts:  ████████░░  83%  — full behavioral transfer

Validated: 83% behavioral fidelity cross-model. 7/7 on domain-agnostic test (bird spotting directory on GPT-5.1). Zero training. Three text files.

Quick start

1. Grab the universal starter pack

Copy universal/MEMORY.md — 10 corrections that benefit every user.

2. Add it to your model's context

Platform Where to put it
Claude Code ~/.claude/CLAUDE.md or memory files
ChatGPT Custom Instructions or paste in first message
Hermes ~/.hermes/MEMORY.md
Any API System message
Any chat Paste at conversation start

3. Use your AI normally

Every time you push back on something, BTP detects it and compiles a new correction. The model gets better with every conversation.

The full agent stack

Layer What it is Who builds it Portable?
Pre-training Model capabilities AI labs Model-locked
RLHF Safety alignment AI labs Model-locked
BTP Behavioral culture You Any model

AI labs build the DNA. BTP builds the culture. DNA is permanent but locked to one model. Culture is maintained but goes anywhere.

Three files, one identity

SOUL.md   → who the AI is (identity, voice, philosophy)
MEMORY.md → how the AI behaves (corrections in BTP format)
USER.md   → who it serves (your preferences, domain, communication style)

Transfer these three files to any model and it acts like your trained assistant from day one.

vs. traditional methods

Fine-tuning BTP
Speed Hours of training Minutes (edit a text file)
Cost GPU compute Zero
Granularity Changes everything Changes one habit
Reversibility Retrain Delete one line
Model lock-in Yes No — works on any model
Verification Eval suites Talk to it and see

The four layers

Each layer teaches the model differently. All are text. All are transferable.

Layer What it does How Compression
L1: Rules Hard boundaries "Never do X" Literal compliance
L2: Pattern Interrupts Override instincts "You'll want X → do Y" First-read effective
L3: Correction Transcripts Simulated experience Show wrong → corrected → right Pattern recognition
L4: Question Chain Compressed runtime Linked questions that drove discovery Recreates search direction
L1 — Rules:              "Never hallucinate."
L2 — Pattern Interrupt:  "YOUR INSTINCT: present 5 options. WHAT WORKS: pick the best one."
L3 — Transcript:         "WRONG: 'Here are some options...' CORRECTION: 'Just pick one.' RIGHT: [picks one]"
L4 — Question Chain:     "1. can it transfer? → 2. why not 100%? → 3. what's the glue?"

How corrections compile

The format that works:

YOUR INSTINCT: [what the model will naturally do]
WHAT ACTUALLY WORKS: [the correct behavior]
TRIGGER: [when this pattern activates]

Why this format: "Don't do X" is noise to a model that's never done X. "You'll want to do X — do Y instead" is actionable from first read. The format predicts the instinct and interrupts it.

Key insight

Culture is portable. Capability isn't.

BTP makes any model behave the way you want. It can't make a weak model stronger. Pre-training defines the ceiling. BTP raises the behavioral floor. Different layers, complementary.

Preferences aren't input — they're output of the correction loop. Nobody fills out a form. They just use it. The corrections accumulate. The system learns what you want by listening to what you push back on.

File structure

btp/
├── README.md                    ← you are here
├── install.sh                   ← one-command full stack setup
├── universal/
│   └── MEMORY.md                ← starter pack (11 universal corrections)
├── compiler/
│   ├── README.md                ← autonomous compilation pipeline
│   ├── btpCompiler.ts           ← Node.js middleware
│   ├── btp-compiler.sh          ← Shell hook (Claude Code)
│   └── session-artifact.sh      ← A=C+Q+O+S session compiler
├── templates/
│   ├── SOUL.md                  ← identity template
│   ├── MEMORY.md                ← correction format template
│   └── USER.md                  ← user profile template
├── spec/
│   └── PROTOCOL.md              ← full technical specification
└── LICENSE

License

MIT

Origin

Discovered during Session 46 of the Verra project (2026-03-25). Started as dispatch plumbing work. An honest self-assessment conversation led to an identity transfer experiment, which revealed that corrections without context are noise. The three-layer format emerged from iterating on what actually changes model behavior — not what describes it.

Built by a Navy CS1 and Claude Opus, testing on GPT-5.1-codex-max via Hermes.

About

Your AI gets better every conversation. No training. Just text files.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages