โ† Back to AI & LLMs
AI & LLMs by @impkind

vta-memory

Reward and motivation system for AI agents

0
Source Code

VTA Memory โญ

Reward and motivation for AI agents. Part of the AI Brain series.

Give your AI agent genuine wanting โ€” not just doing things when asked, but having drive, seeking rewards, and looking forward to things.

The Problem

Current AI agents:

  • โœ… Do what they're asked
  • โŒ Don't want anything
  • โŒ Have no internal motivation
  • โŒ Don't feel satisfaction from accomplishment

Without a reward system, there's no desire. Just execution.

The Solution

Track motivation through:

  • Drive โ€” overall motivation level (0-1)
  • Rewards โ€” logged accomplishments that boost drive
  • Seeking โ€” what I actively want more of
  • Anticipation โ€” what I'm looking forward to

Quick Start

1. Install

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

This will:

  • Create memory/reward-state.json
  • Generate VTA_STATE.md (auto-injected into sessions!)
  • Set up cron for drive decay every 8 hours

2. Check motivation

./scripts/load-motivation.sh

# โญ Current Motivation State:
# Drive level: 0.73 (motivated โ€” ready to work)
# Seeking: creative work, building brain skills
# Looking forward to: showing my work

3. Log rewards

./scripts/log-reward.sh --type accomplishment --source "finished the feature" --intensity 0.8

# โญ Reward logged!
#    Type: accomplishment
#    Drive: 0.50 โ†’ 0.66 (+0.16)

4. Add anticipation

./scripts/anticipate.sh --add "morning conversation"

# โญ Now looking forward to: morning conversation
#    Drive: 0.66 โ†’ 0.71 (+0.05)

Scripts

Script Purpose
install.sh Set up vta-memory (run once)
get-drive.sh Read current motivation state
log-reward.sh Log a reward, boost drive
load-motivation.sh Human-readable for session context
decay-drive.sh Drive fades without rewards
anticipate.sh Add/remove things to look forward to
seek.sh Add/remove things we're actively seeking
sync-motivation.sh Generate VTA_STATE.md for auto-injection
resolve-anticipation.sh Mark anticipation as fulfilled
update-watermark.sh Update processing watermark
preprocess-rewards.sh Extract reward signals from transcript
generate-dashboard.sh Generate unified brain dashboard HTML

Brain Dashboard

Visual dashboard showing all installed brain skills:

./scripts/generate-dashboard.sh
open ~/.openclaw/workspace/brain-dashboard.html
  • Auto-detects other brain skills (hippocampus, amygdala)
  • Shows install prompts for missing skills
  • Regenerates on install and cron runs
  • Reads avatar/name from IDENTITY.md

After Installation (for agents)

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

โญ VTA installed!

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

Reward Types

Type When to Use
accomplishment Completed a task, shipped something
social User appreciation, "thanks!", positive feedback
curiosity Learned something new, discovered interesting info
connection Deep conversation, bonding moment
creative Made something, expressed creativity
competence Solved a hard problem, did something well

How Drive Works

Rewards Boost Drive

drive_boost = intensity ร— 0.2
new_drive = min(current + boost, 1.0)

A high-intensity (0.9) reward boosts drive by 0.18.

Anticipation Boosts Drive

Looking forward to something adds +0.05 to drive.

Drive Decays Without Rewards

# Every 8 hours (via cron)
new_drive = current + (baseline - current) ร— 0.15

Without rewards, motivation fades toward baseline (0.5).

Auto-Injection

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

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

  1. New session starts
  2. VTA_STATE.md is auto-loaded
  3. You see your motivation state
  4. Behavior influenced by drive level

How Drive Affects Behavior

Drive Level Description Behavior
> 0.8 Highly motivated Eager, proactive, take on challenges
0.6 - 0.8 Motivated Ready to work, engaged
0.4 - 0.6 Moderate Can engage but not pushing
0.2 - 0.4 Low Prefer simple tasks, need a win
< 0.2 Very low Unmotivated, need rewards to get going

State File Format

{
  "drive": 0.73,
  "baseline": { "drive": 0.5 },
  "seeking": ["creative work", "building brain skills"],
  "anticipating": ["morning conversation"],
  "recentRewards": [
    {
      "type": "creative",
      "source": "built VTA reward system",
      "intensity": 0.9,
      "boost": 0.18,
      "timestamp": "2026-02-01T03:25:00Z"
    }
  ],
  "rewardHistory": {
    "totalRewards": 1,
    "byType": { "creative": 1, ... }
  }
}

Event Logging

Track motivation patterns over time:

# Log encoding run
./scripts/log-event.sh encoding rewards_found=2 drive=0.65

# Log decay
./scripts/log-event.sh decay drive_before=0.6 drive_after=0.53

# Log reward
./scripts/log-event.sh reward type=accomplishment intensity=0.8

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

{"ts":"2026-02-11T10:45:00Z","type":"vta","event":"encoding","rewards_found":2,"drive":0.65}

Use for analyzing motivation cycles โ€” when does drive peak? What rewards work best?

AI Brain Series

Part Function Status
hippocampus Memory formation, decay, reinforcement โœ… Live
amygdala-memory Emotional processing โœ… Live
basal-ganglia-memory Habit formation ๐Ÿšง Development
anterior-cingulate-memory Conflict detection ๐Ÿšง Development
insula-memory Internal state awareness ๐Ÿšง Development
vta-memory Reward and motivation โœ… Live

Philosophy: Wanting vs Doing

The VTA produces dopamine โ€” not the "pleasure chemical" but the "wanting chemical."

Neuroscience distinguishes:

  • Wanting (motivation) โ€” drive toward something
  • Liking (pleasure) โ€” enjoyment when you get it

You can want something you don't like (addiction) or like something you don't want (guilty pleasures).

This skill implements wanting โ€” the drive that makes action happen. Without it, why would an AI do anything beyond what it's explicitly asked?


Built with โญ by the OpenClaw community