Three Tier Memory - Aristidis Vasilopoulos
Problem
"LLM-based agentic coding assistants lack persistent memory: they lose coherence across sessions, forget project conventions, and repeat known mistakes."
Main Contribution
A three-component infrastructure developed during construction of a 108,000-line C# system:
- Hot-memory constitution (660 lines, always loaded) - Encodes conventions, retrieval hooks, and orchestration protocols
- Specialized agents - 19 domain-expert agents (9,300 lines total) invoked per task
- Cold-memory knowledge base - 34 on-demand specification documents (~16,250 lines) queried via MCP retrieval server
Research Scope
- Development sessions analyzed: 283
- Case studies: 4 observational studies showing how codified context prevents failures and maintains consistency
- Scale: Applied to a distributed system with significant complexity
Key Innovation
The framework enables codified context to propagate across sessions, addressing scalability challenges for multi-agent projects that previous manifest file approaches couldn't handle.
Publication Details
- Submitted: February 24, 2026
- License: Creative Commons BY 4.0
- Open-source companion repository with DOI