Memory and always-loaded context

Context Most AI tools have amnesia. Yours does not.

Two kinds of memory

There is a context window, the agent's short-term memory for the current conversation. Fixed size. Old messages drop off. Start a new chat and it's empty. There is long-term memory, what the agent remembers across sessions. No size limit. It can hold years of context. Long-term memory splits into two buckets that you should never confuse: Always-loaded context, general behavior. Loaded at the start of every session. How the agent works, no matter the task. ("Always send progress updates while you're working." "Never wait silently, keep me posted.") Skills, specific task behavior. Loaded only when called. Recipes for recurring jobs. ("How to format my emails." "How to summarize a transcript and extract action items." "How to draft a proposal.") If a rule is about how the agent operates in general, it goes in always-loaded context. If a rule is about how to do a specific task, it's a skill. You never edit either of these by hand. You tell the agent what you want and it writes the rule in the right place. The simple analogy: always-loaded context is what the agent memorizes, its name, your name, your address, the rules of the road. Skills are the reference books on its shelf. It doesn't memorize a cookbook; it grabs the cookbook when it's time to cook. Treat alwaysloaded context as the agent's brain, small, focused, always with it, and skills as the library it reaches for when it needs them.

The bloat problem

Always-loaded context costs tokens every session. If it gets bloated, the model starts ignoring rules, there's too much in front of it for any single rule to feel important. Symptoms: The agent confirms a rule is there but doesn't follow it. It forgets things mid-session that it has rules about.

Long sessions feel slower and dumber than short ones. Fix: run this once a week. Audit your always-loaded context. For each rule:

1. MOVE TO SKILLS: if it's about how to do a specific task, extract it into a new or existing skill so it only loads when needed.
2. CUT REDUNDANCY: if the same idea appears twice, keep one canonical version.
3. COMPRESS: rewrite what remains in fewer tokens. Tables beat paragraphs. Bullets beat prose.
4. REPORT: show me what you moved, cut, and compressed. Goal: minimize always-loaded context so the rules that remain have maximum weight in long sessions.

Run this and you will be shocked how much sludge accumulates in a month.

Project folders and build logs

For bigger pieces of work, ask the agent to set up a projects folder with a subfolder per project. Each subfolder gets a plan file and a build log. Add a rule to always-loaded context: "before working on any project, read its plan and build log first." That plus good skills = no memory problem, even on multi-month builds. What goes in always-loaded context, what goes in skills, and how to keep both clean.