Every session, you start from zero.
You open a new chat and the AI has no idea who you are, what you work on, how you approach a brief, or what you've already figured out. So you explain. You paste context. You describe your setup. And next session, you do it again.
This is the hidden cost most people don't talk about. It's not the AI's output that's the problem. It's the setup tax you pay every single time.
The Reset Is Eating Your Time
Chat windows have no memory across sessions. When the window fills up or you close it, everything you built with the AI disappears. The context you carefully loaded, the preferences you established, the history of decisions you worked through. All of it gone when you close the window.
For someone using AI occasionally, that's annoying. For someone using it as a serious part of their work process, it's a real drag on what the tool can actually do for you.
The fix isn't to keep one endless conversation open and hope it doesn't hit a context limit. The fix is to stop relying on the chat history at all.
What Permanent Memory Actually Looks Like
The idea is simple: build a knowledge base outside the chat. A set of plain text files that hold everything important about you, your work, your processes, and your preferences. Files any AI can read at the start of any session.
No proprietary format. No cloud sync with unknown privacy terms. Just Markdown files on your own machine.
Andrej Karpathy has written about this kind of locally-hosted LLM wiki, the idea being that your knowledge lives in a form that any capable AI can instantly parse. You write it once, in plain language, and it becomes the permanent context layer that travels with you across every tool.
Paired with Obsidian, which stores everything locally as Markdown files, this becomes something more than a note-taking system. It becomes the memory your AI pulls from instead of relying on chat history.
The Obsidian Layer
Obsidian stores everything as plain Markdown files on your hard drive. No proprietary database, no account required to access your own writing. That matters for AI use because the files are readable by any AI tool that can access your local file system.
The setup is not complicated. You create vaults in Obsidian for the information you want AI to know: your positioning, your clients, your active projects, your preferences, your processes. Each note is a plain text file. When you start a new AI session, you reference the relevant files instead of typing out context from memory.
What changes is that the knowledge compounds. Every time you figure something out, capture a preference, or document a process, you add it to the vault. That work doesn't disappear when you close a chat window. It stays, and it's immediately available the next time you need it.
What This Looks Like in Practice
Instead of starting a session by explaining who you are and what you need, you point the AI to the relevant files. A note on your work style. A note on the project in progress. A note on the client you're drafting for.
The AI reads them at the start of the conversation and works from that foundation instead of a blank slate.
You're not re-explaining your tone preferences every time you draft something. You're not re-establishing your positioning before every piece of content. You wrote that once, it lives in Obsidian, and any session can start from it.
The other advantage: zero privacy risk with your most sensitive knowledge. The files live on your machine. You're not feeding your client list, your pricing structure, or your creative process into a cloud sync you don't control.
How to Actually Build This
You need a tool that can reach the files on your own machine. A coding assistant like Claude Code or Codex, or VS Code with an AI assistant set up. Anything that can read and write files where your notes live, not just chat in a browser tab.
Then it's three steps.
First, install Obsidian and create a vault. This is just a folder on your hard drive where everything gets stored as plain Markdown.
Second, tell Claude that from now on it should use that Obsidian vault as your knowledge base. The place it reads from at the start of a session and writes to whenever something is worth keeping.
Third, hand it the LLM wiki. Paste Karpathy's link and let Claude read it. That one document is the blueprint for how the whole thing should be structured, and Claude will set up the vault for you off the back of it.
From there you just feed it. Drop in a transcript, paste a block of text, point it at a document, and Claude files it into your second brain in a form it can read back later. The more you put in, the more it knows the next time you sit down to work.
The One Shift That Changes Everything
This isn't really about tools. It's about where the knowledge lives.
Right now most people's AI setup treats every conversation as the full unit of value. The chat is the product. That's why losing the chat feels like losing the work.
Once you move the knowledge out of the chat and into a local system you own, the chat becomes what it actually is: a work session. The knowledge you built doesn't depend on that session staying open. It lives somewhere else, and the next session picks up from the same foundation.
That's what makes AI useful in a sustained way rather than impressive in a single demo.
The creative work itself still requires you. The taste, the ear, the instinct for what belongs in a piece and what doesn't. None of that lives in a Markdown file. But the context around the work, the systems, the preferences, the accumulated decisions about how you operate, all of that can be captured, stored, and instantly available to whatever AI you're working with next.
You write it once. It works forever.
