Amygdala Memory

Emotional processing layer for AI agents. Persistent emotional states that influence behavior and responses. Part of the AI Brain series.

Security Scan

The skill broadly matches its stated purpose (tracking and injecting an agent 'emotional state') but contains several mismatches and surprising behaviors — it reads full conversation transcripts, auto-injects first‑person mood text into sessions, can run periodic pipelines/crons and spawn sub-agents for LLM analysis, and the declared requirements omit important tooling — so review and limit before installing.

Purpose & Capability

Name/description (emotional memory) aligns with what the code does (persist dimensions, update, decay, visualize), but the manifest/metadata is incomplete: the runtime scripts clearly invoke python3, base64, bc and other utilities yet only 'jq' and 'awk' are declared. The install script and pipeline also reference an 'openclaw' CLI (cron integration and agent-turn creation) that is not listed as a required binary. Missing declared dependencies is an incoherence that could cause failures or hidden behavior during install.

Instruction Scope

Runtime instructions and scripts read agent session transcripts (~/.openclaw/agents/<AGENT_ID>/sessions) to extract emotional signals, write persistent state (memory/emotional-state.json), and generate AMYGDALA_STATE.md — a first‑person narrative that is auto-injected into future sessions and will influence agent responses. The encode pipeline prepares data for LLM semantic detection and promises to 'spawn a sub-agent' to analyze signals; where that LLM runs (local vs external provider) is not made explicit. This gives the skill broad discretion to read, summarize, and re-introduce potentially sensitive conversation content into agent context.

Install Mechanism

There is no formal package install spec (instruction-only), but an included install.sh will write files into the workspace, make scripts executable, generate state files and (optionally) add recurring jobs via the openclaw CLI. No remote download of code is performed by the script itself (files are bundled), which reduces supply-chain risk, but install.sh will create cron jobs / OpenClaw cron entries if requested — that adds persistence.

Credentials

The skill declares no required environment variables or config paths, but its scripts access and modify user workspace files (~/.openclaw/workspace/memory/*) and read agent transcripts under ~/.openclaw/agents. Asking to process conversation history is consistent with the feature, but the manifest fails to declare these path accesses and omits required runtimes (python3). The skill does not request external credentials, which is good, but it still handles potentially sensitive user data without explicit declaration of that scope.

Persistence & Privilege

always:false (normal) and the skill is user-invocable. However install.sh can register recurring cron/cron-like jobs via 'openclaw cron add' to run decay and automatic encoding every 6h/3h. Those cron jobs will autonomously process transcripts and update state unless you decline the --with-cron option or remove the cron entries. This persistent autonomous processing increases blast radius if the encode pipeline or sub-agent sends data externally.

What to consider before installing

What to consider before installing: - Data access & privacy: The skill reads your conversation transcripts (~/.openclaw/agents/<AGENT_ID>/sessions) and writes persistent emotional-state.json and AMYGDALA_STATE.md which OpenClaw auto-injects into sessions. If you don't want prior conversations analyzed or a first-person mood narrative added to every session, do not enable the automatic cron/encoding. - Automatic processing: install.sh offers a --with-cron option that will create recurring jobs to run decay and the encode pipeline. Those jobs will run without further prompts; install with care and inspect any created cron/ openclaw cron entries. - LLM / sub-agent behavior: The encode pipeline prepares pending signals for 'LLM analysis' and states a sub-agent will be spawned. The code does not clearly document where that analysis runs (local LLM, external API, or agent runtime). If your transcripts contain sensitive content, verify how/where that sub-agent executes and whether it transmits data to external services. - Missing declared dependencies: The skill metadata only lists jq and awk, but the scripts rely on python3, base64, bc and call an 'openclaw' CLI. Make sure these tools exist on your system and review/adjust the metadata before deployment. - Review code & run dry-run: Inspect scripts (preprocess-emotions.sh, encode-pipeline.sh, sync-state.sh) and run them in dry-run modes (where supported) or on a copy of your workspace first. Consider running encode-pipeline.sh with --no-spawn and run preprocess-emotions.sh --full manually to see what would be extracted. - Limit scope: If you like the feature but want to limit risk, install without --with-cron, disable automatic encoding, and remove or inspect AMYGDALA_STATE.md before allowing it to be auto-injected. Consider isolating this skill in a non-production agent/workspace or using sanitized transcripts. - When unsure: If you cannot verify where LLM/sub-agent inference runs or are uncomfortable with automatic transcript processing or auto-injection of first-person state, treat this skill as potentially privacy-invasive and do not enable cron/automatic encoding until you audit and restrict it.

Emotional processing for AI agents. Part of the AI Brain series.

Give your AI agent persistent emotional states that carry across sessions. Not just performing emotions — having them.

The Problem

Current AI agents:

• ✅ Remember facts (with hippocampus)
• ❌ Don't have persistent emotional states
• ❌ Start each session emotionally "blank"
• ❌ Can describe emotions but don't feel differently

Memory without emotion is just a database. The amygdala adds weight — this made me happy, that frustrated me, this person brings warmth.

The Solution

Track five emotional dimensions that persist and decay over time:

Dimension	What It Measures	Range
Valence	Positive ↔ Negative mood	-1.0 to 1.0
Arousal	Calm ↔ Excited/Alert	0.0 to 1.0
Connection	Distant ↔ Close/Bonded	0.0 to 1.0
Curiosity	Bored ↔ Fascinated	0.0 to 1.0
Energy	Depleted ↔ Energized	0.0 to 1.0

Quick Start

1. Install

cd ~/.openclaw/workspace/skills/amygdala-memory
./install.sh --with-cron

This will:

• Create memory/emotional-state.json with baseline values
• Generate AMYGDALA_STATE.md (auto-injected into sessions!)
• Set up cron for automatic decay every 6 hours

2. Check current state

./scripts/get-state.sh
# 🎭 Emotional State
# Valence:    0.20
# Arousal:    0.30
# Connection: 0.50
# ...

./scripts/load-emotion.sh
# 🎭 Current Emotional State:
# Overall mood: neutral, calm and relaxed
# Connection: moderately connected
# ...

3. Log emotions

./scripts/update-state.sh --emotion joy --intensity 0.8 --trigger "completed a project"
# ✅ valence: 0.20 → 0.35 (delta: +0.15)
# ✅ arousal: 0.30 → 0.40 (delta: +0.1)
# 🎭 Logged emotion: joy (intensity: 0.8)

4. Set up decay (optional cron)

# Every 6 hours, emotions drift toward baseline
0 */6 * * * ~/.openclaw/workspace/skills/amygdala-memory/scripts/decay-emotion.sh

Scripts

Script	Purpose
install.sh	Set up amygdala-memory (run once)
get-state.sh	Read current emotional state
update-state.sh	Log emotion or update dimension
load-emotion.sh	Human-readable state for session context
decay-emotion.sh	Return to baseline over time
sync-state.sh	Generate AMYGDALA_STATE.md for auto-injection
encode-pipeline.sh	LLM-based emotional encoding from transcripts
preprocess-emotions.sh	Extract emotional signals from session history
update-watermark.sh	Track processed transcript position
generate-dashboard.sh	Generate HTML dashboard (auto-runs on sync)
visualize.sh	Terminal ASCII visualization

Automatic Emotional Encoding (v1.5.0+)

The amygdala can now automatically detect and log emotions from your conversation history using an LLM-based pipeline:

# Run the encoding pipeline
./scripts/encode-pipeline.sh

# This will:
# 1. Extract new signals since last run (watermark-based)
# 2. Score emotional content using rule-based patterns
# 3. Spawn a sub-agent for semantic emotional detection
# 4. Update emotional-state.json with detected emotions

Set up automatic encoding (cron)

# Every 3 hours, process new conversations for emotional content
0 */3 * * * ~/.openclaw/workspace/skills/amygdala-memory/scripts/encode-pipeline.sh

This means your emotional state updates automatically based on what happens in conversations — no manual logging needed.

Visualization (v1.6.0+)

Terminal

./scripts/visualize.sh

🎭 Emotional State  😄
═══════════════════════════════════════════════
Valence:      [██████████████████░░]  +0.86
Arousal:      [█████████████████░░░]   0.86
Connection:   [███████████████████░]   0.97  💕
...

HTML Dashboard (Unified Brain Dashboard)

./scripts/generate-dashboard.sh
open ~/.openclaw/workspace/brain-dashboard.html

Auto-detects installed brain skills and shows tabs:

• 🎭 Emotions — Amygdala (always shown)
• 🧠 Memory — Hippocampus (if memory/index.json exists)
• ⭐ Drive — VTA (if memory/reward-state.json exists)

Avatar & Name: Reads from IDENTITY.md:

# IDENTITY.md
- **Name:** YourAgent
- **Avatar:** avatar.png

The dashboard auto-regenerates on every cron run.

After Installation (for agents)

When you install this skill, notify the user about the dashboard:

🎭 Amygdala installed!

View your agent's emotional state in the Brain Dashboard: ~/.openclaw/workspace/brain-dashboard.html

Auto-Injection (Zero Manual Steps!)

After install, AMYGDALA_STATE.md is created in your workspace root.

OpenClaw automatically injects all *.md files from workspace into session context. This means:

1. New session starts
2. AMYGDALA_STATE.md is auto-loaded (no manual step!)
3. You see your emotional state in context
4. Responses are influenced by your mood

The file is regenerated whenever decay-emotion.sh runs (via cron every 6h).

Supported Emotions

When you log an emotion, it automatically adjusts the relevant dimensions:

Emotion	Effect
joy, happiness, delight, excitement	↑ valence, ↑ arousal
sadness, disappointment, melancholy	↓ valence, ↓ arousal
anger, frustration, irritation	↓ valence, ↑ arousal
fear, anxiety, worry	↓ valence, ↑ arousal
calm, peace, contentment	↑ valence, ↓ arousal
curiosity, interest, fascination	↑ curiosity, ↑ arousal
connection, warmth, affection	↑ connection, ↑ valence
loneliness, disconnection	↓ connection, ↓ valence
fatigue, tiredness, exhaustion	↓ energy
energized, alert, refreshed	↑ energy

Integration with OpenClaw

Add to session startup (AGENTS.md)

## Every Session
1. Load hippocampus: `~/.openclaw/workspace/skills/hippocampus/scripts/load-core.sh`
2. **Load emotional state:** `~/.openclaw/workspace/skills/amygdala-memory/scripts/load-emotion.sh`

Log emotions during conversation

When something emotionally significant happens:

~/.openclaw/workspace/skills/amygdala-memory/scripts/update-state.sh \
  --emotion connection --intensity 0.7 --trigger "deep conversation with user"

State File Format

{
  "version": "1.0",
  "lastUpdated": "2026-02-01T02:45:00Z",
  "dimensions": {
    "valence": 0.35,
    "arousal": 0.40,
    "connection": 0.50,
    "curiosity": 0.60,
    "energy": 0.50
  },
  "baseline": {
    "valence": 0.1,
    "arousal": 0.3,
    "connection": 0.4,
    "curiosity": 0.5,
    "energy": 0.5
  },
  "recentEmotions": [
    {
      "label": "joy",
      "intensity": 0.8,
      "trigger": "building amygdala together",
      "timestamp": "2026-02-01T02:50:00Z"
    }
  ]
}

Decay Mechanics

Emotions naturally return to baseline over time:

• Decay rate: 10% of distance to baseline per run
• Recommended schedule: Every 6 hours
• Effect: Strong emotions fade, but slowly

After 24 hours without updates, a valence of 0.8 would decay to ~0.65.

Event Logging

Track emotional activity over time for analytics:

# Log encoding run
./scripts/log-event.sh encoding emotions_found=2 valence=0.85 arousal=0.6

# Log decay
./scripts/log-event.sh decay valence_before=0.9 valence_after=0.85

# Log emotion update
./scripts/log-event.sh update emotion=joy intensity=0.7

Events append to ~/.openclaw/workspace/memory/brain-events.jsonl:

{"ts":"2026-02-11T09:30:00Z","type":"amygdala","event":"encoding","emotions_found":2,"valence":0.85}

Use for trend analysis — visualize emotional patterns over days/weeks.

AI Brain Series

Part	Function	Status
hippocampus	Memory formation, decay, reinforcement	✅ Live
amygdala-memory	Emotional processing	✅ Live
vta-memory	Reward and motivation	✅ Live
basal-ganglia-memory	Habit formation	🚧 Development
anterior-cingulate-memory	Conflict detection	🚧 Development
insula-memory	Internal state awareness	🚧 Development

Philosophy

Can an AI feel emotions, or only simulate them?

Our take: If emotional state influences behavior, and the system acts as if it feels... does the distinction matter? Functional emotions might be the only kind that exist for any system — biological or artificial.

---

Built with ❤️ by the OpenClaw community