Why Your AI Forgets Your Project

You open a new session and your assistant, the same one that wrote half your codebase, asks what the project does. So you explain. Again. You paste the same context, re-describe the same architecture, restate the same goals. By the time it is oriented, you have spent twenty minutes and a chunk of your patience getting back to where you were yesterday.

It happens in Cursor when you open a new chat, in Claude Code after /clear, and in Codex at the start of a run. The tools differ. The morning is the same.

Why every session starts cold

Underneath, these tools work the same way. The model only sees what is in its context window right now: your messages, the files it has opened, and whatever instructions were loaded at the start. Cursor's docs on rules say the models behind it do not keep memory between completions. Anthropic's docs say each Claude Code session begins with a fresh context window.

A new chat opens an empty window. Running /clear empties the one you were in. A long session that fills its window gets summarised or cut short, and a summary keeps the gist and tends to drop the reasoning; the Claude Code version of that is in Claude Code Loses Context.

The code survives all of it, because the code is files. What you understood about the project lived in the conversation, and the conversation is the part that goes.

What people try, and where each one stops

Longer chats. The first instinct is to never close the window. Then the window fills, the tool summarises or you start over, and you are back at the cold start with a longer recap to write.

Rules files. Every tool has a file it reads up front. CLAUDE.md in Claude Code. Rules in .cursor/rules, or an AGENTS.md, in Cursor. AGENTS.md in Codex, which reads it before doing any work. Project knowledge in Lovable, which it reads with every message. Have one. It is the right home for how to build: stack, conventions, things never to touch. It makes a poor home for where the project is going, and it is small on purpose: Lovable caps project knowledge at 10,000 characters. Files describe. They don't decide. Nothing checks the finished work against them, and they never reorder the backlog when reality changes.

Memory add-ons. These write notes about your corrections and preferences and load them next time; Claude Code now has one built in. They do save you from repeating yourself about tabs and test commands. Memory recalls what happened. It doesn't hold a plan. A memory of everything you did is not a plan for what's next.

What the session actually needs

Look at what your morning recap actually says. It is rarely history. It is: here is what we are building, and for whom. Here is the task in front of you, and what done looks like. Here is what we decided last week, why, and the option we ruled out. Here is what comes after this.

A cold session does not need to know what happened. It needs to know what to do next, and what not to undo.

That is a plan and a set of decisions, kept current. It changes every day, which is why a hand-kept file goes stale in a busy week. It belongs to the project, not to one tool's chat or rules format. Changing tools makes that obvious, as switching AI coding tools without starting over covers.

A useful test: could a session that has never seen the project pick up the next task and get it right? That is a cold handoff, and it is a better bar than any amount of recall.

One way to run it

This is the job PAPI does, inside the tool you already build in. Work runs in cycles of a few tasks. Each task gets a build handoff with the goal, what is out of scope and how to check it is done, written so an assistant with no history on the project can run it. Decisions keep their reasoning and a confidence level. A changed decision is superseded, not overwritten, so the next session sees what you think now and why. Every build files a report, and the next plan reads it.

Because the project connects to whichever tool you open, Cursor on Monday and Claude Code on Tuesday read the same plan.

The honest version

PAPI does not make the context window bigger, and it does not write code. Every session still starts cold. What changes is what it reads first: the plan and the decisions, instead of your recap. Review stays with you, task by task.

If your sessions are one task each, nothing from last week matters this week, and a short rules file covers the rest, stay there. A plan outside the chat earns its place once the project has outgrown what you can hold in your head. See that plan on a demo project, no signup needed.