Articles
Thinking out loud about AI infrastructure.
Design notes, cognitive architecture deep dives, and lessons from building Continuum Engine.
Activation
Your brain has a retrieval algorithm. Psychologists spent forty years reverse-engineering it. Your search bar has about a third of it, and nobody seems to have noticed.
Forgetting Is a Forecast
Every time your memory lets something go, it is placing a bet about tomorrow - using odds it has been quietly collecting your entire life.
What Brings a Memory Back
History decides how deep a memory has sunk. Whatever is already in your head right now decides what gets pulled back up, and it has an odd side effect: the more you know about something, the harder each fact about it is to reach.
The Rival Theory
There is a second serious theory of how thinking works, and it disagrees with the first one almost everywhere. The places they independently agreed amount to a set of guardrails for machine memory.
Activation in Practice
Every dial in the machine, what it actually does, and what measuring it changed.
The tools do the math. The model does the judgment.
Agent Sterling - a .NET code quality reviewer on Microsoft Agent Framework that separates deterministic Roslyn analysis from LLM editorial judgment. Four tools, one system prompt, a workflow executor, and the line between agent and service.
MU-TH-UR 6000: Building a durable AI agent in .NET.
Temporal workflows, M.E.AI, pgvector, and .NET Aspire - a paint-by-numbers guide to a durable AI agent with PDF extraction, document storage, vector search, and child workflows.
Checkpointing LLM chains with Temporal
Multi-step LLM inference chains are structurally fragile. Temporal workflow durability turns retry-from-scratch into replay-from-checkpoint.
Choosing local LLMs for a two-model pipeline
Model selection, quantization, and LM Studio configuration for a two-model inference pipeline on 32 GB of VRAM.
No article matches that search, so try fewer words or another topic.