Project Memory for LLM Systems
A concise argument for entity graphs, specs, and memory-like structure around LLM operating systems.
An LLM operating system needs more than chat history. It needs a durable map of entities, relationships, constraints, and actions. Without that map, every task begins with rediscovery.
A project memory layer should tell the model what the important things are, how they connect, and which surface is authoritative. That is a different job from autocomplete or summarization.
Why graphs help
Entity graphs reduce ambiguity. Instead of asking a model to scan a pile of prose and infer relationships, the project can expose names, links, states, and constraints directly. The agent can then spend less context on navigation and more context on the actual task.
This is the reason specdog leans on .dog files and compiled .dag graphs. The graph is not decoration; it is the operational map.
The site as memory surface
A public site can participate in that memory layer if it is cleanly structured. The homepage names the product family. The about page explains the purpose. The handbook teaches the workflow. The agent page gives runtime instructions. The blog gives background. The sitemap and llms.txt make those pages easier to discover.