Why your AI assistant forgets what you were building (and how to fix it)

You open a new Claude Code session. You type your first message. And then, before you can get any real work done, you spend the next ten minutes re-explaining your project.

The tech stack. The naming conventions. Why you’re using that particular state management approach. The constraint you ran into two weeks ago that ruled out the obvious solution. The decision you made on Friday about how to structure the API layer. All of it, gone. You’re not picking up where you left off. You’re starting over with a collaborator who has no idea who you are.

This is context debt, and it compounds every single session.

The frustrating part is that none of this knowledge actually disappeared. It’s in your git history, your comments, your commit messages. The reasoning exists somewhere. But your AI assistant can’t see any of it when a new session opens. Each conversation is a clean slate.

This happens because of how large language models work at a fundamental level. There’s no persistent memory between sessions. The model has no access to what you discussed yesterday. It doesn’t read your codebase before you start talking. It only knows what you put in the current context window, and that window opens empty every time. The model isn’t forgetting, exactly. It never knew. You have to rebuild the shared understanding from scratch on every session.

The result is a real pattern: the longer you work on a project, the more context overhead each session carries. Early on, you can get straight to the code. Six months in, you’re spending the first part of every session reconstructing architectural history before you can ask a useful question.

The fix is giving your AI assistant a structured briefing it can actually read at session start. Not a README, which is written for humans and covers the happy path. A session briefing: what was last worked on, what decisions were made and why, what the next step is, and what constraints are in play right now.

KeepGoing’s MCP server does exactly this. You run npx @keepgoingdev/mcp-server once to set it up, and it connects to Claude Code via the Model Context Protocol. When you start a session, Claude can call the get_briefing tool to pull your last checkpoint before you’ve typed a single message. The architectural constraints you documented last Thursday are there. The decision you made about the database schema is there. The file you were halfway through is there.

Instead of ten minutes rebuilding shared context, you paste one line and get to work.

If you’re already using Claude Code regularly, check whether MCP is configured in your project. The setup takes about two minutes, and the difference shows up immediately on the next session where you would have otherwise started cold.