Projects vs conversations

Working with Claude on Projects

Chapter 1  ·  Projects vs Conversations — Persistent Context and What Claude Remembers

There's a fundamental difference between chatting with Claude and working with Claude. A chat is a one-off exchange. Working with Claude — on a codebase, a course, a research project, a business workflow — benefits from something more structured: a Project. This chapter explains what Projects are, what they change, and what Claude actually remembers across sessions (and what it doesn't).

Conversations vs. Projects — The Core Difference

A conversation
  • Exists for the duration of one session
  • No memory carries over when you start a new chat
  • You re-explain context every time: your stack, your goals, your preferences
  • Claude treats you as a stranger at the start of each new message thread
  • Great for one-off questions, quick lookups, exploratory chats
A Project
  • A persistent workspace that spans many sessions over days, weeks, or months
  • Has a system prompt (instructions) that Claude reads at the start of every session
  • Supports a memory system — files that Claude can read and update across sessions
  • Can include uploaded files, documents, and code as persistent context
  • Claude behaves like a long-term collaborator who knows your project

The practical effect: in a Project, you invest in context once and it compounds. Every session starts from a shared foundation rather than a blank slate.

What Claude Actually Remembers — and What It Doesn't

This is the most important thing to understand about working with Claude long-term. The word "memory" is used loosely — the reality is more precise and more useful once you understand it.

📋
System prompt / Project instructions
Written once, loaded into every session automatically. This is where you put permanent rules, context, and preferences. Claude reads this before your first message every time.
Persists always
🗂️
Memory files (MEMORY.md + individual files)
Markdown files Claude writes and reads. Store facts about you, your preferences, project decisions, and reference pointers. Claude reads these at the start of sessions and updates them when asked or when it learns something important.
Persists always
📁
Uploaded project files
Documents, PDFs, code files you attach to the Project. Available to Claude across sessions. Good for reference material, schemas, style guides, existing code.
Persists always
💬
Current session conversation
Everything said in the current session — Claude reads the whole thread on every message. But this resets when you start a new session. It is not automatically saved to memory.
This session only
🔮
Implicit knowledge / "things Claude learned"
Claude does not learn from conversations in the way a human does. If you tell Claude something important in a session and it isn't written to a memory file, it's gone next session. There is no background learning.
Does not persist
The most common misconception
"I told Claude this last week." If it wasn't written to a memory file or the system prompt, Claude has no record of it. The conversation happened — but Claude's next session starts fresh. This is why the memory system matters: it's explicit, not magical.

What's in Claude's Context Window During a Project Session

Every time you send a message in a Project, Claude reads all of this before generating a response:

Instructions
Memory
Project files
Session history
Your message
free
System prompt / instructions Memory files Uploaded project files This session's conversation Your current message

The total context window is 200,000 tokens — enormous for most projects. The key insight is that the persistent layers (instructions, memory, files) are loaded fresh each session, so they don't accumulate session-by-session. Only the conversation history grows within a session, and it resets when you start a new one.

What Persists Between Sessions — Full Table

WhatPersists?How
Project instructions (system prompt) Always Written in Project settings; loaded automatically every session
Memory files (user, feedback, project, reference) Always Written by Claude or you; read by Claude at session start
Uploaded documents and files Always Attached to the Project; available every session
Preferences you stated in a conversation Never Unless Claude wrote them to a memory file — tell Claude to remember them
Decisions made mid-session Never Ask Claude to save important decisions to the project memory file
Code Claude wrote in a session Only if saved Code in the filesystem persists; code only in the chat does not
The conversation thread itself Viewable, not loaded You can scroll back and read it; Claude does not re-read past sessions
Claude's "learning" from interactions Never Claude does not update its model weights from conversations

When to Use a Project vs. a Conversation

Use a Project
Ongoing development work

Building an app, writing a course, maintaining a server. You'll have many sessions and context accumulates — a Project makes every session smarter than the last.

Use a Project
You have strong preferences

Editor shortcuts, code style, preferred tools, tone for writing. Write these into memory once and never re-explain them again.

Use a Project
Reference material you use repeatedly

A database schema, API docs, a style guide, a project spec. Upload once to the Project; Claude can read it any session.

Use a Project
You want consistent behaviour across sessions

Same response style, same assumptions, same constraints. The system prompt enforces this automatically — you don't have to re-state it.

Use a conversation
One-off questions

"What does this error mean?" "How do I do X in Python?" No context needed across sessions — a conversation is lighter and faster.

Use a conversation
You want a fresh, unbiased perspective

If you want Claude without any context shaping its answer — no prior decisions, no project framing — a new conversation gives you a clean slate.

The Mental Model: Claude as a Long-Term Collaborator

The most useful way to think about a Project is as working with a colleague who reads a briefing document before every meeting. That briefing contains everything they need to know about you and the project — but they don't remember the actual conversation from last week unless someone took notes.

Your job as the project owner is:

  • Write the briefing — the system prompt: who you are, what the project is, how you like to work
  • Keep the notes — the memory files: decisions made, preferences discovered, important facts
  • Maintain the filing cabinet — uploaded files: reference material Claude should always have access to

When those three layers are well-maintained, every new session feels like continuing a conversation — because Claude walks in already knowing the context.

In this project — right now
This course is itself running inside a well-configured Project. The system prompt tells Claude about the website, the file structure, and the course format. Memory files store your learning profiles, course outlines, and preferences (like always using vi, never nano). Uploaded files aren't used here, but the other two layers are — and you've seen the results across dozens of sessions.
Next — Chapter 2: The Memory System
A deep dive into the four memory types — user, feedback, project, and reference — how they work, when to create each one, and how to write memory entries that are actually useful across sessions rather than just archives of what happened.