Self-Improvement System

The Agent Gateway includes a built-in self-improvement engine that converts conversation history into reusable learnings and skills.

How it works

1. Facet extraction — after conversations, the system analyzes your session history and extracts patterns, preferences, and domain knowledge
2. Synthesis — extracted facets are combined into reusable skills and knowledge fragments
3. Application — synthesized learnings are applied to future conversations, making the agent more effective over time

This is all wallet-scoped — your learnings are private and never shared with other wallets.

For the full architecture, verification gates, and research basis, see Self-Improvement & Research Validation.

Configuration

CLI

rickydata wallet settings

API

# Check self-improvement status
curl -s https://agents.rickydata.org/wallet/self-improvement/status \
  -H "Authorization: Bearer mcpwt_..."

# Manually trigger a self-improvement run
curl -s -X POST https://agents.rickydata.org/wallet/self-improvement/trigger \
  -H "Authorization: Bearer mcpwt_..."

# View history
curl -s https://agents.rickydata.org/wallet/self-improvement/history \
  -H "Authorization: Bearer mcpwt_..."

Settings

Control via PUT /wallet/settings:

Setting	Values	Description
autoImprove	true / false	Enable/disable automatic runs
improveSchedule	after_each, daily, weekly, biweekly	When to run

Schedule options

Schedule	When it runs	Best for
after_each	After every conversation	Heavy users who want maximum learning
daily	Once per day	Balanced approach
weekly	Once per week	Light users
biweekly	Every two weeks	Minimal overhead

Cost

Self-improvement runs consume wallet balance:

• Each run uses your BYOK API key to call Claude for analysis
• The 10% platform markup applies to LLM costs
• Runs are subject to your wallet's budget limits

You can set a retention budget to cap how much the system spends on self-improvement:

rickydata wallet settings
# Set retention budget in the settings

What gets extracted

The facet extraction pipeline looks for:

• Preferences — how you like responses formatted, level of detail, communication style
• Domain knowledge — topics you frequently discuss, terminology you use
• Tool patterns — which tools you use most, common argument patterns
• Workflow patterns — multi-step operations you repeat

Monitoring

Status check

curl -s https://agents.rickydata.org/wallet/self-improvement/status \
  -H "Authorization: Bearer mcpwt_..." | jq .

Returns:

• Whether auto-improve is enabled
• Current schedule
• Last run timestamp
• Number of extracted facets
• Number of synthesized skills

History

curl -s https://agents.rickydata.org/wallet/self-improvement/history \
  -H "Authorization: Bearer mcpwt_..." | jq '.runs[-1]'

Shows recent runs with:

• Run timestamp
• Sessions analyzed
• Facets extracted
• Skills synthesized
• Cost of the run

Privacy

• All self-improvement data is wallet-scoped and encrypted
• Facets and skills are stored on the Agent Gateway's encrypted disk
• No cross-wallet data sharing occurs
• Delete your data by deleting your sessions and triggering a cleanup

Next steps

• Spending Policy — budget controls
• Sessions — session management
• Self-Improvement & Research Validation - how the system verifies private skills