The cost of starting over
AI chat memory helps, but it doesn't create a portable project record. Keep decisions, instructions, and work context in a durable layer.
Open a fresh chat and your assistant may already know a little about you. ChatGPT can save memories and reference past chats. Claude can build memory across chats. That’s good.
The work still crosses their boundaries. Change vendors, open a coding client, or move to another machine, and chat memory doesn’t give you a shared project record. The exact decision from three days ago, or the work someone else needs to read, can disappear into the gaps.
That’s the cost of starting over. Not total amnesia. Rebuilding a brief that should already exist: your stack, the decision, the reason behind it, the work already done.
The usual workaround is a bigger prompt. You keep a doc of “context to paste,” and you paste it. It works until it doesn’t. The doc goes stale, it lives on one machine, and it only covers the assistant you copied it into.
Keep the project context outside any one conversation or vendor, in one place, and let compatible assistants read it directly. When the work moves forward, write the new piece back. The next session can begin from the same record. Here’s how to share that context between AI tools.
A knowledge layer is a small, durable set of things worth remembering: who you are, how you work, what you’re building. It is a work record, not a pile of chat transcripts to tend by hand.
MCP makes it practical. Connect a client to the layer once and it can read the background when a new session begins. Ask it to note a decision and it lands somewhere you’ll find again, not in a thread you’ll lose by Friday. The wiring takes about a minute — here’s the walkthrough for Claude.
You’re still in charge. The layer doesn’t act on its own, and it won’t nag you to feed it. It keeps an authored record readable when the client changes.
The next question is what to put in it, and the answer is less than you’d think. Four kinds of context earn their place; the rest is noise that makes search worse.
vtriv is built around this idea. If the retyping has been getting to you, give it a try: get started.