Create and train AI learning plugins with AgentDB's 9 reinforcement learning algorithms. Includes Decision Transformer, Q-Learning, SARSA, Actor-Critic, and more. Use when building self-learning agents, implementing RL, or optimizing agent behavior through experience.
npxskills add ruvnet/claude-flow--skill "AgentDB Learning Plugins"Loading…
Provides access to 9 reinforcement learning algorithms via AgentDB's plugin system. Create, train, and deploy learning plugins for autonomous agents that improve through experience. Includes offline RL (Decision Transformer), value-based learning (Q-Learning), policy gradients (Actor-Critic), and advanced techniques.
Performance: Train models 10-100x faster with WASM-accelerated neural inference.
# Interactive wizard
npx agentdb@latest create-plugin
# Use specific template
npx agentdb@latest create-plugin -t decision-transformer -n my-agent
# Preview without creating
npx agentdb@latest create-plugin -t q-learning --dry-run
# Custom output directory
npx agentdb@latest create-plugin -t actor-critic -o .$plugins
# Show all plugin templates
npx agentdb@latest list-templates
# Available templates:
# - decision-transformer (sequence modeling RL - recommended)
# - q-learning (value-based learning)
# - sarsa (on-policy TD learning)
# - actor-critic (policy gradient with baseline)
# - curiosity-driven (exploration-based)
# List installed plugins
npx agentdb@latest list-plugins
# Get plugin information
npx agentdb@latest plugin-info my-agent
# Shows: algorithm, configuration, training status
import { createAgentDBAdapter } from 'agentic-flow$reasoningbank';
// Initialize with learning enabled
const adapter = await createAgentDBAdapter({
dbPath: '.agentdb$learning.db',
enableLearning: true, // Enable learning plugins
enableReasoning: true,
cacheSize: 1000,
});
// Store training experience
await adapter.insertPattern({
id: '',
type: 'experience',
domain: 'game-playing',
pattern_data: JSON.stringify({
embedding: await computeEmbedding('state-action-reward'),
pattern: {
state: [0.1, 0.2, 0.3],
action: 2,
reward: 1.0,
next_state: [0.15, 0.25, 0.35],
done: false
}
}),
confidence: 0.9,
usage_count: 1,
success_count: 1,
created_at: Date.now(),
last_used: Date.now(),
});
// Train learning model
const metrics = await adapter.train({
epochs: 50,
batchSize: 32,
});
console.log('Training Loss:', metrics.loss);
console.log('Duration:', metrics.duration, 'ms');
Type: Offline Reinforcement Learning Best For: Learning from logged experiences, imitation learning Strengths: No online interaction needed, stable training
npx agentdb@latest create-plugin -t decision-transformer -n dt-agent
Use Cases:
Configuration:
{
"algorithm": "decision-transformer",
"model_size": "base",
"context_length": 20,
"embed_dim": 128,
"n_heads": 8,
"n_layers": 6
}
Type: Value-Based RL (Off-Policy) Best For: Discrete action spaces, sample efficiency Strengths: Proven, simple, works well for small$medium problems
npx agentdb@latest create-plugin -t q-learning -n q-agent
Use Cases:
Configuration:
{
"algorithm": "q-learning",
"learning_rate": 0.001,
"gamma": 0.99,
"epsilon": 0.1,
"epsilon_decay": 0.995
}
Type: Value-Based RL (On-Policy) Best For: Safe exploration, risk-sensitive tasks Strengths: More conservative than Q-Learning, better for safety
npx agentdb@latest create-plugin -t sarsa -n sarsa-agent
Use Cases:
Configuration:
{
"algorithm": "sarsa",
"learning_rate": 0.001,
"gamma": 0.99,
"epsilon": 0.1
}
Type: Policy Gradient with Value Baseline Best For: Continuous actions, variance reduction Strengths: Stable, works for continuous$discrete actions
npx agentdb@latest create-plugin -t actor-critic -n ac-agent
Use Cases:
Configuration:
{
"algorithm": "actor-critic",
"actor_lr": 0.001,
"critic_lr": 0.002,
"gamma": 0.99,
"entropy_coef": 0.01
}
Type: Query-Based Learning Best For: Label-efficient learning, human-in-the-loop Strengths: Minimizes labeling cost, focuses on uncertain samples
Use Cases:
Type: Robustness Enhancement Best For: Safety, robustness to perturbations Strengths: Improves model robustness, adversarial defense
Use Cases:
Type: Progressive Difficulty Training Best For: Complex tasks, faster convergence Strengths: Stable learning, faster convergence on hard tasks
Use Cases:
Type: Distributed Learning Best For: Privacy, distributed data Strengths: Privacy-preserving, scalable
Use Cases:
Type: Transfer Learning Best For: Related tasks, knowledge sharing Strengths: Faster learning on new tasks, better generalization
Use Cases:
// Store experiences during agent execution
for (let i = 0; i < numEpisodes; i++) {
const episode = runEpisode();
for (const step of episode.steps) {
await adapter.insertPattern({
id: '',
type: 'experience',
domain: 'task-domain',
pattern_data: JSON.stringify({
embedding: await computeEmbedding(JSON.stringify(step)),
pattern: {
state: step.state,
action: step.action,
reward: step.reward,
next_state: step.next_state,
done: step.done
}
}),
confidence: step.reward > 0 ? 0.9 : 0.5,
usage_count: 1,
success_count: step.reward > 0 ? 1 : 0,
created_at: Date.now(),
last_used: Date.now(),
});
}
}
// Train on collected experiences
const trainingMetrics = await adapter.train({
epochs: 100,
batchSize: 64,
learningRate: 0.001,
validationSplit: 0.2,
});
console.log('Training Metrics:', trainingMetrics);
// {
// loss: 0.023,
// valLoss: 0.028,
// duration: 1523,
// epochs: 100
// }
// Retrieve similar successful experiences
const testQuery = await computeEmbedding(JSON.stringify(testState));
const result = await adapter.retrieveWithReasoning(testQuery, {
domain: 'task-domain',
k: 10,
synthesizeContext: true,
});
// Evaluate action quality
const suggestedAction = result.memories[0].pattern.action;
const confidence = result.memories[0].similarity;
console.log('Suggested Action:', suggestedAction);
console.log('Confidence:', confidence);
// Store experiences in buffer
const replayBuffer = [];
// Sample random batch for training
const batch = sampleRandomBatch(replayBuffer, batchSize: 32);
// Train on batch
await adapter.train({
data: batch,
epochs: 1,
batchSize: 32,
});
// Store experiences with priority (TD error)
await adapter.insertPattern({
// ... standard fields
confidence: tdError, // Use TD error as confidence$priority
// ...
});
// Retrieve high-priority experiences
const highPriority = await adapter.retrieveWithReasoning(queryEmbedding, {
domain: 'task-domain',
k: 32,
minConfidence: 0.7, // Only high TD-error experiences
});
// Collect experiences from multiple agents
for (const agent of agents) {
const experience = await agent.step();
await adapter.insertPattern({
// ... store experience with agent ID
domain: `multi-agent/${agent.id}`,
});
}
// Train shared model
await adapter.train({
epochs: 50,
batchSize: 64,
});
// Collect batch of experiences
const experiences = collectBatch(size: 1000);
// Batch insert (500x faster)
for (const exp of experiences) {
await adapter.insertPattern({ /* ... */ });
}
// Train on batch
await adapter.train({
epochs: 10,
batchSize: 128, // Larger batch for efficiency
});
// Train incrementally as new data arrives
setInterval(async () => {
const newExperiences = getNewExperiences();
if (newExperiences.length > 100) {
await adapter.train({
epochs: 5,
batchSize: 32,
});
}
}, 60000); // Every minute
Combine learning with reasoning for better performance:
// Train learning model
await adapter.train({ epochs: 50, batchSize: 32 });
// Use reasoning agents for inference
const result = await adapter.retrieveWithReasoning(queryEmbedding, {
domain: 'decision-making',
k: 10,
useMMR: true, // Diverse experiences
synthesizeContext: true, // Rich context
optimizeMemory: true, // Consolidate patterns
});
// Make decision based on learned experiences + reasoning
const decision = result.context.suggestedAction;
const confidence = result.memories[0].similarity;
# Create plugin
npx agentdb@latest create-plugin -t decision-transformer -n my-plugin
# List plugins
npx agentdb@latest list-plugins
# Get plugin info
npx agentdb@latest plugin-info my-plugin
# List templates
npx agentdb@latest list-templates
// Reduce learning rate
await adapter.train({
epochs: 100,
batchSize: 32,
learningRate: 0.0001, // Lower learning rate
});
// Use validation split
await adapter.train({
epochs: 50,
batchSize: 64,
validationSplit: 0.2, // 20% validation
});
// Enable memory optimization
await adapter.retrieveWithReasoning(queryEmbedding, {
optimizeMemory: true, // Consolidate, reduce overfitting
});
# Enable quantization for faster inference
# Use binary quantization (32x faster)
npx agentdb@latest mcpCategory: Machine Learning / Reinforcement Learning Difficulty: Intermediate to Advanced Estimated Time: 30-60 minutes
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Set up and use 1Password CLI (op). Use when installing the CLI, enabling desktop app integration, signing in (single or multi-account), or reading/injecting/running secrets via op.
CLI to manage emails via IMAP/SMTP. Use `himalaya` to list, read, write, reply, forward, search, and organize emails from the terminal. Supports multiple accounts and message composition with MML (MIME Meta Language).
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Set up and use 1Password CLI (op). Use when installing the CLI, enabling desktop app integration, signing in (single or multi-account), or reading/injecting/running secrets via op.
CLI to manage emails via IMAP/SMTP. Use `himalaya` to list, read, write, reply, forward, search, and organize emails from the terminal. Supports multiple accounts and message composition with MML (MIME Meta Language).