Writing
Richard Wollyce, on building Ulpia: what was decided, what it cost, and what is still wrong.
-
Why In-Process Memory Beats Resident Daemons: Lessons from the AI Memory Frontier
Exploring the architectural tradeoffs in AI agent memory: why in-process SQLite FTS5 beats resident server daemons, how filesystem boundaries prevent multi-tenant leaks, and what is coming next in Ulpia.
-
From Prototype to Production: 5 Architectural Upgrades in Ulpia's RAG & Multi-Agent Engine
Five architectural upgrades turning naive RAG into a production engine: small-to-big section windowing, constitutional critique ledgers, cross-domain query decomposition, offline triad grading, and 4-bit AWQ student distillation under 2 GB RAM.
-
Our system doesn't need a model to do the retrieval
The first video on the channel: why I decided to build Ulpia, why AI memory should be local, and why retrieval here does not go through a model.
-
The confidence floor now scales with your corpus
Why one threshold cannot serve a base of ten notes and a base of a thousand, every term in the formula spelled out, and the three measurements the change had to pass before it shipped.