AI Compass / Quick reads
Memory belongs in files, not in the chat
What you tell an AI only in the chat is gone by the next chat. What is in a file stays.
In depth: Memory in three tiers – what Claude knows about you and your projects
Jack Roberts builds memory in three tiers: a global instruction about you under 200 words, a folder per work area with a short CLAUDE.md and notes on decisions, and an archive into which every session writes a dated summary at the end. His yardstick: a new chat answers just as well right away. In Claude Code the global instruction is ~/.claude/CLAUDE.md. In addition, according to the docs, Claude Code itself creates notes about your preferences and corrections and loads them into every session: the automatic memory (docs checked).
Tool and alternative
Local Markdown files are enough; Obsidian for reading is optional. Pinecone from the video is a cloud database and is not needed for one person. The finished “Memory OS” is in Jack’s community. “10x” in the title is unsupported.
Sources
- Jack Roberts, “This Memory System just 10x’d Claude Code”, 3 May 2026 (3:11-3:43, 4:49, 6:23-7:29)
- Jack Roberts, “Karpathy’s Skills just changed Everything (Claude Code)”, 28 April 2026 (12:21)
- Claude Code docs Memory, retrieved 17 October 2026
Advertising in the source: Firecrawl affiliate link, Glaido and community (3 May), short links suspected to be affiliate links (28 April).
Want to try this with your team? That’s what the AI Lab is for. See the AI Lab