How to make Claude stop forgetting everything

If you've ever closed a Claude session and felt like you lost everything you built — you're not imagining it.

Every session starts from zero. Your architecture decisions, your conventions, the edge case you debugged for an hour. Gone when the window closes.

Teams lose 30–40 minutes per session just re-explaining what they already knew yesterday. That's a full workday per week. Every week.

The fix isn't a smarter model. It's giving it a memory system.

Here's the exact 3-layer architecture:

Layer 1 — Teach it who you are

CLAUDE.md is the first thing Claude reads every session. Most people treat it like a settings file. The builders getting 10x output treat it like onboarding a new hire who never forgets.

What belongs in it: your architecture decisions, naming conventions, workflow preferences, and hard boundaries — what to never do.

Pair it with an auto-memory directory that persists Claude's observations between sessions:

~/.claude/projects/<project-hash>/memory/
├── MEMORY.md        # routing document — keep under 200 lines
├── debugging.md     # solutions to recurring problems
├── patterns.md      # confirmed conventions
├── architecture.md  # key decisions
└── preferences.md   # your workflow

The session rhythm that makes this work: orient → work → persist. Every session is a deposit. They compound.

Layer 2 — Give it a searchable brain

An Obsidian vault wired to Claude via MCP (Model Context Protocol) turns your notes into a live knowledge graph Claude can query at runtime.

Two MCP servers make this work:

• smart-connections — semantic search over your entire vault. Claude finds relevant notes even without knowing the exact title.

• qmd — structured queries and metadata operations.

The highest-leverage change you can make: name your notes as claims, not categories.

❌ memory-systems.md ✅ memory graphs beat giant memory files.md

When Claude searches your vault, the titles alone tell it whether something is relevant — before it reads a single line.

Layer 3 — Feed the brain automatically

Most knowledge doesn't start as text. It starts as a YouTube video, a voice memo, a client meeting.

brain-ingest closes the gap. One command:

bash

brain-ingest "https://youtube.com/watch?v=..." --apply

It downloads, transcribes locally, extracts structured claims and frameworks, and drops a note into your Obsidian inbox. From a single 90-minute talk it typically pulls 12–18 distinct claims, 3–5 frameworks, and 5–8 actionable techniques.

That's the material that compounds when Claude retrieves it six weeks later on a completely different problem.

The setup checklist:

• Create CLAUDE.md in your project root

• Enable auto-memory in Claude Code settings

• Set up Obsidian vault with structured folders

• Install smart-connections MCP — pip install smart-connections-mcp

• Install qmd MCP — npx -y @tobilu/qmd mcp

• Run brain-ingest on your last 3 most valuable recordings

• Commit to the rhythm: orient → work → persist

The first session after setup feels different. Not because Claude got smarter. Because it stopped forgetting.

→ Full guide with MCP config, vault structure, and every config file