Class 3: AI Worker Scheduling, Memory & Projects

Class 3

AI Worker Scheduling, Memory & Projects#

Organizing an AI worker's world — isolated projects, retrieved context (RAG), memory, instructions, and work that runs on a schedule.

By Ghulam Qasim Husnain · Sun, Oct 4, 2026 projectscontextRAGmemoryscheduled tasksinstructions

Once a worker can act, the next job is organizing its world: wall off each domain in a project, feed it the right context, let memory and instructions do their separate jobs, and hand repeat work to scheduled tasks.

From Sir Zia’s class — Zia Tutor AI, The AI Agent Factory (FDE track).

Why it matters

Capability without structure gets messy fast — memories bleed across domains, context balloons, and you retype the same setup every session. Projects, context and scheduling are how an architect keeps a worker organized and reliable.

Proactiveness & scheduled tasks

Proactive — Claude can start a conversation or take a step itself, not only when you prompt it.
Scheduled tasks run work automatically at a set date and time — set up by asking Claude or by hand. Each run starts a brand-new session, and you tell it up front which connectors, skills and plugins to use. Example: a daily task that pulls a lesson from a source you’ve wired up with a connector.
You’re told a run started in one of two ways right now: a push notification or an email.

Correction Because every scheduled run is a fresh session, it only sees what persists across sessions — project context, connectors, memory, skills. A PDF or page you dropped into one chat won’t be there. So to use a PDF or website as a recurring source, put it in the project’s context or behind a connector, not a one-off upload.

Projects — isolation and boundaries

Project A walled-off workspace. It keeps everything separate — sessions (chat threads), memory, even scheduled tasks — so you can work across many domains without their context or memory bleeding together.

A project is about organization, isolation, boundaries, and dividing work into domains. Its real gift: your work stays within those boundaries.

Correction “All sessions are mixed and can’t be filtered” is only true without projects — then everything shares one space and memory blends across domains. Projects exist precisely to separate sessions, memory and instructions per domain, so you’re not stuck with a mix: use a project per domain.

Context — the worker’s reference shelf

Context A file, a set of files, or data the worker consults when it needs to — like a book you open only to the pages relevant to your question.

Context is two-way: you can upload files, and the worker can create and hand files back.

My note Does an unrelated file (say a birthday-planner) sitting in the project pollute a chat? Mostly no — context is fetched by semantic search / RAG, so only the relevant pieces are pulled in. Still, keep a project’s context focused; less clutter means better retrieval.

Memory vs. instructions — the distinction that matters

Key distinction Instructions go with every prompt — always-on rules the worker can’t miss. Memory is recalled on demand: the worker uses tools to look it up when relevant. Write instructions that are well-defined and specific.

Project instructions in practice

Instead of clicking /Zia Tutor AI every time I want to learn, I can put that in the project’s instructions so it happens on its own:

Example Project instruction: “Whenever I start learning, use the Zia Tutor AI connector and skill.” Other useful ones: “use practical examples from my field (SEO / accounting),” “make images and keep the text brief.”

A principle to design by

Remember For every AI worker you design, write down its value, its cost, and its priority — then trade off between them deliberately.

My plan — top things I’ll build with Claude

Zia Tutor AI — for learning.
Social-media management — e.g. LinkedIn posts.
Email marketing.
Job work — e.g. a data analyst’s day-to-day.
Development & interview prep.

Each would get its own project (isolated memory, instructions, sessions), and may use connectors, plugins and skills to do the work.

Questions I’m still exploring

Any downside to running many scheduled tasks? What’s the limit, and the minimum interval?
Can two tasks — two sessions — run at the same time?
How do shared “common” instructions interact with per-project instructions?

Test yourself

recall questions

Why can’t a scheduled task use a PDF you dropped into a different chat?
→ Each run is a new session; it only sees what persists across sessions — project context, connectors, memory, skills.
What does a project isolate?
→ Sessions, memory and tasks — plus its own instructions — per domain.
Instructions vs memory?
→ Instructions ride with every prompt; memory is recalled on demand via tools.
Why doesn’t an unrelated context file usually interfere?
→ RAG / semantic search retrieves only the relevant pieces.