Self-Improving AI Agents for the Enterprise
Governed self-improvement for enterprise AI agents
AI agents can now remember previous interactions, develop reusable skills, and improve through experience. But what happens when an agent learns the wrong lesson, relies on outdated information, or changes its behaviour without sufficient oversight?
In this remote session, Sathish Palanisamy, a former Googler will introduce a governed memory and improvement layer built around evidence, review, measurement, auditability, and rollback for autonomous agents.
Through practical examples, we will examine how agent experiences can be converted into durable lessons, how conflicting or incorrect knowledge can be corrected, and how organizations can retain control over what their agents learn.
The session will cover:
What “self-improving” actually means for an AI agent
The difference between agent memory and measurable improvement
What is Hermes and how it creates and reuses skills
How DejaDB records evidence and governs learned knowledge
Detecting repeated tool failures and contradictory memories
Reviewing and approving agent-generated improvements
Measuring whether a new lesson improves performance
Rolling back or superseding harmful and outdated knowledge
Architectural considerations for enterprise agent systems
About the Presenter
Sathish Palanisamy is the Founder and CEO of MindGryd Software, an agentic AI company focused on helping organizations move from AI experimentation to reliable, production-scale agent systems.
Sathish brings extensive technology and product experience from companies including Google, Amazon, and Yahoo. His work spans large-scale software systems, AI platforms, developer infrastructure, and the practical challenges involved in deploying intelligent systems inside real business workflows.
At MindGryd, he is focused on building AI agents that combine autonomy with human oversight, governance, and measurable business outcomes.
Who Should Attend?
This session is designed for:
AI and agent-platform developers
Engineering and technology leaders
Enterprise architects
Product managers building AI-powered workflows
Developers evaluating agent memory and learning systems
Researchers interested in reliable and governed AI agents
Basic familiarity with AI agents and tool calling will be helpful, but prior experience with Hermes or DejaDB is not required.
Event Format
This is a live remote session featuring:
Architectural explanations
Practical self-improvement examples
A Hermes and DejaDB comparison
Demonstrations of learning, correction, and rollback
Live Q&A
Join us to explore how agents can improve through experience—without losing visibility or control over what they learn.