Synix
An open-source agent-memory build system with DAG execution, content-addressed artifacts, incremental rebuilds, and source-level provenance.
Mark Lubin · AI Infrastructure & Agent Systems
Dependable systems for long-running agents.
I build infrastructure for AI agents: runtimes and orchestration, memory and context systems, evaluation, and the backend systems that keep them reliable.
Selected work
LENS
A benchmark for long-horizon agent coherence under context accumulation and drift, comparing eight memory architectures over 120-episode sequences.
Compress, Divide, Renew
An architectural framework for coherence in persistent agent systems through lossy compression, timescale partitioning, and renewal.
Talk: What Memory Is (And Isn't)
Slides from a talk on what memory is — and isn't — in agent systems.
Technical and technology writing
- Agency Is Not Autonomy
- Coherence in Persistent Adaptive Systems
- Papers Worth Reading If You're Thinking About Agent Memory
- Hopping Context Windows
- Agent Skills Are Frozen Intent
- Agentic Transactions: A Conceptual Framework for Agent-Mediated Collaboration and Negotiation
- Experiential Policy Memory
- Toward Emergent Memory Architecture for AI Systems
- I Spent Weeks Mapping the World's Most Effective Startup Incubator. It's Not What You Think.
About
I spent ten years at Amazon building distributed systems across payments, Prime Video, and AWS Marketplace. From January 2025 through March 2026, I focused on independent research and open-source engineering for persistent agent systems. That work produced Synix, LENS, and CDR. I now work as a Staff Software Engineer at Kinelo, building shared infrastructure for agent runtimes, sessions, events, and operations.
Contact
For professional inquiries: mark@synix.dev