Emotion Skill

Emotions for AI agents — built from memories, shaped by relationships, always changing.

Emoclaw trains a lightweight model on your agent's identity and conversation history, producing a persistent emotional state that evolves with every interaction. Emotions decay naturally between sessions, shift based on who's talking and what's being said, and gradually recalibrate as your agent grows. The result is injected into the system prompt as an [EMOTIONAL STATE] block, giving your AI a felt sense of its own inner life.

How it works

1. 1. Bootstrap — extract.py reads your agent's identity/memory files. label.py scores each passage via the Claude API (opt-in). train builds a small neural net from those scores. One-time setup.
2. Encode — Each incoming message is turned into a 384-dim vector by a frozen MiniLM sentence encoder. No fine-tuning, no network call — runs from a local cache.
3. Feel — The encoding + context (who's talking, what channel, previous emotion) flows through a GRU and MLP head, outputting an N-dimensional emotion vector (0-1 per dimension). The GRU hidden state persists across sessions — this is the "emotional residue" that carries forward mood.
4. Decay — Between sessions, each dimension drifts back toward its baseline at a configurable half-life (fast for arousal, slow for safety/groundedness). Time apart = cooling off.
5. Inject — The emotion vector is formatted as an [EMOTIONAL STATE] block and inserted into the agent's system prompt, giving the AI a felt sense of its own inner state.

Model is ~2MB, runs on CPU, adds <50ms per message. Network access is only used during bootstrap (opt-in).

Quick Reference

Setup

Quick Setup

CODEBLOCK0

This copies the bundled emotion_model engine to your project root, creates a venv, installs the package, and copies the config template. Then edit emoclaw.yaml to customize for your agent.

Manual Setup

If you prefer to set up manually:

1. Install the package

CODEBLOCK1

Required: Python 3.10+, PyTorch, sentence-transformers, PyYAML.

2. Copy and customize the config

CODEBLOCK2

Edit emoclaw.yaml to set:

• - name — your agent's name
• INLINECODE25 — emotional dimensions with baselines and decay rates
• INLINECODE26 — map of relationship names to embedding indices
• INLINECODE27 — communication channels your agent uses
• INLINECODE28 — absence-based desire growth (can be disabled)
• INLINECODE29 — cpu recommended (MPS has issues with sentence-transformers)

See references/config-reference.md for the full schema.

3. Bootstrap (new agent)

If starting from scratch with identity/memory files:

CODEBLOCK3

Or run the full pipeline:

CODEBLOCK4

4. Verify

CODEBLOCK5

Usage

Option A: Daemon (Recommended)

The daemon loads the model once and listens on a Unix socket, avoiding the ~2s sentence-transformer load time per message.

CODEBLOCK6

Option B: Direct Python Import

CODEBLOCK7

Option C: One-shot State Injection

For system prompt injection without the daemon:

CODEBLOCK8

This reads the persisted state, applies time-based decay, and outputs the [EMOTIONAL STATE] block.

Integration

System Prompt Injection

Add the output block to your system prompt. The block format:

CODEBLOCK9

Daemon Protocol

Send JSON over the Unix socket:

CODEBLOCK10

Special commands:
CODEBLOCK11

Heartbeat Integration

The emotional state decays over time and needs to be refreshed at each session start. Add this entry to your HEARTBEAT.md:

CODEBLOCK12

Or call the daemon / inject_state script from your heartbeat/cron:

CODEBLOCK13

Important: Without heartbeat integration, the emotional state block will go stale between sessions. The inject_state script applies time-based decay and outputs the current state — it must be called at least once per session.

Architecture

The model processes each message through this pipeline:

CODEBLOCK14

The GRU hidden state persists across sessions — this is the "emotional residue" that carries forward mood, context, and relational memory.

See references/architecture.md for full details.

Security & Privacy

Data Flow

1. 1. Extraction (scripts/extract.py) reads markdown files listed in emoclaw.yaml → bootstrap.source_files and bootstrap.memory_patterns. These are configurable and default to identity/memory files within the repo. Extracted passages are written to emotion_model/data/extracted_passages.jsonl.

1. 2. Redaction — Before writing, extracted text is passed through configurable regex patterns (bootstrap.redact_patterns) that replace API keys, tokens, passwords, and other secrets with [REDACTED]. Default patterns cover Anthropic keys, GitHub PATs, bearer tokens, SSH keys, and generic key=value credentials. Add custom patterns in emoclaw.yaml.

1. 3. Labeling (scripts/label.py) — opt-in only. Sends extracted passages to the Anthropic API for emotional scoring. Requires both ANTHROPIC_API_KEY and explicit user consent (interactive prompt before any API call). Use --yes to skip the prompt for automation. Use --dry-run to preview without any network calls.

1. 4. Training runs entirely locally. No data leaves the machine during prepare_dataset or train.

1. 5. Inference runs entirely locally. The daemon and inject_state script make no network calls.

Network Access

Network access is optional and limited to a single script:

The sentence-transformers encoder downloads model weights on first use (from Hugging Face). After that, it runs from cache with no network needed.

File Permissions

The daemon socket is created with permissions 0o660 (owner + group read/write) and cleaned up on shutdown. The socket path is configurable in emoclaw.yaml → paths.socket_path.

Path Validation

INLINECODE66 validates that every file path resolves to within the repository root before reading. Symlink chains and ../ sequences that would escape the repo boundary are rejected. This prevents a misconfigured source_files or memory_patterns from reading arbitrary files.

Configuring Redaction

Add or modify patterns in emoclaw.yaml:

CODEBLOCK15

Set redact_patterns: [] to disable redaction entirely (not recommended).

Isolation Recommendations

• - Run the bootstrap pipeline (extract → label → train) in an isolated environment or review the source file list before running
• Audit bootstrap.source_files and bootstrap.memory_patterns in your emoclaw.yaml to ensure only intended files are included
• Review emotion_model/data/extracted_passages.jsonl before running label.py to confirm no sensitive content will be sent externally
• The daemon should run under the same user as your agent process — avoid running as root

Configuration

All configuration lives in emoclaw.yaml. The package falls back to built-in defaults if no YAML is found.

Config search order:

1. 1. EMOCLAW_CONFIG environment variable
2. INLINECODE79 (project root)
3. INLINECODE80

Key sections:

• - dimensions — name, labels, baseline, decay half-life, loss weight
• INLINECODE82 — known senders with embedding indices
• INLINECODE83 — communication channels (determines context vector size)
• INLINECODE84 — absence-based desire modulation
• INLINECODE85 — architecture hyperparameters
• INLINECODE86 — training hyperparameters
• INLINECODE87 — self-calibrating baseline drift (opt-in)

See references/config-reference.md for the complete schema.

Bootstrap Pipeline

Step 1: Extract Passages

INLINECODE89 reads identity and memory files, splitting them into labeled passages:

CODEBLOCK16

Source files are configured in emoclaw.yaml → bootstrap.source_files and bootstrap.memory_patterns.

Step 2: Auto-Label

INLINECODE93 uses the Claude API to score each passage on every emotion dimension:

CODEBLOCK17

Each passage gets a 0.0-1.0 score per dimension plus a natural language summary.

Step 3: Prepare & Train

CODEBLOCK18

Retraining

To add new training data:

1. 1. Add entries to emotion_model/data/ in JSONL format:

   {"text": "message text", "labels": {"valence": 0.7, "arousal": 0.4, ...}}
   

1. 2. Re-run the preparation and training:

Incremental Retraining

The training script saves a rich checkpoint (training_checkpoint.pt) that preserves the full optimizer state, learning rate schedule, and early stopping counter. To continue training from where you left off:

CODEBLOCK21

This is a true continuation — optimizer momentum, cosine annealing position, and patience counter all pick up exactly where they stopped.

Growth Model

As the AI accumulates real conversation data:

1. 1. Passive collection — Log messages + model predictions
2. Correction events — When emotion feels wrong, log the correction
3. Periodic retraining — Incorporate new data, retrain
4. Baseline adjustment — Baselines may shift as the AI develops

The system is designed to grow with the AI, not remain static.

Resources

• - references/architecture.md — Model architecture deep-dive
• INLINECODE97 — Full YAML config schema
• INLINECODE98 — Emotion dimension documentation
• INLINECODE99 — Baseline, decay, and self-calibration tuning
• INLINECODE100 — Version upgrade guide
• INLINECODE101 — Template config for new AIs
• INLINECODE102 — Generic summary templates
• INLINECODE103 — Example personality-specific templates
• INLINECODE104 — Bundled emotion_model Python package (copied to project root by setup.py)