Context Kit

Problem

AI coding agents forget everything between sessions.

Solution

One file. Optionally two.

.ai/MEMORY.md      ← always
.ai/DESIGN.md      ← only if architecture is non-trivial
.ai/archive/       ← historical snapshots

Setup

Create .ai/MEMORY.md:

## Active Task

## Current Context

## Key Decisions

## Code Patterns

## Gotchas & Lessons

## Next Steps

Create .ai/DESIGN.md only if needed:

## Architecture

How It Works

1. Session start — Read .ai/MEMORY.md (and .ai/DESIGN.md if it exists)
2. During work — Update after meaningful changes
3. Session end — Keep under ~150 lines

When It Grows Too Large

1. Copy full .ai/MEMORY.md → .ai/archive/MEMORY_YYYY-MM-DD.md
2. Compress .ai/MEMORY.md down to ~80 lines
3. Keep: active task, key decisions, current gotchas
4. Drop: completed tasks, obsolete context

For Your Project

Works with Claude Code, Gemini CLI, and Cursor

This convention is just markdown files in your repo, so you can use it with any AI coding tool that can read project instructions and edit files.

Add the following snippet to your tool’s project instruction file:

• Claude Code CLI: CLAUDE.md (repo root)
• Qwen: QWEN.md (repo root)
• Cursor: .cursor/rules/ai-memory.md (recommended) or your existing Cursor rules/instructions file
• Gemini CLI: your Gemini CLI project instruction file (use whatever file your setup already loads for repo-level guidance; if you don’t have one yet, create a repo-root instruction file and point Gemini CLI at it)

## Memory

Read `.ai/MEMORY.md` at session start.
Read `.ai/DESIGN.md` if it exists.
Update after meaningful changes.
Keep under ~150 lines.

### When MEMORY.md approaches 150 lines:
1. Copy full content to `.ai/archive/MEMORY_YYYY-MM-DD.md`
2. Compress `.ai/MEMORY.md` to ~80 lines
3. Keep: active task, key decisions, current gotchas
4. Drop: completed tasks, obsolete context

No Tool Required

The AI reads and writes markdown files natively. No CLI, no package, no friction. Just files and a convention.