Memory Bank

Your Claude Code sessions die too fast. You waste tokens re-explaining your project every time.

memory-bank fixes both problems: persistent memory + 3-5x longer sessions.

Install · Token Savings · Features · Quick Start · Architecture · Examples · Contributing

The Two Problems Nobody Has Solved Together

Problem 1: Claude forgets everything between sessions.

Every time you open Claude Code, you get: *"So, what are we working on today?"*
You've spent 6 sessions building an auth system. Claude remembers none of it.

Problem 2: Sessions hit context limits way too fast.

You spend 1,200-3,000 tokens re-explaining your project before doing any
real work. That's 10-25% of your session burned on warm-up. Over 10 sessions,
that's 12,000-30,000 tokens wasted just getting Claude back to baseline.

One Skill Fixes Both

Install memory-bank. Now you get this:

No re-explaining. No lost context. No wasted tokens.

Token Savings (Verified with tiktoken)

Real numbers. Measured, not estimated.

Where the Savings Come From

*All numbers measured using tiktoken (cl100k_base) on our example MEMORY.md files.*

Install

Then add the auto-setup to your CLAUDE.md (or ~/.claude/CLAUDE.md for all projects):

That's it. Claude now remembers everything, forever.

Features

This is not a simple note-taker. It's an adaptive memory engine.

Layered Memory Architecture

Layer 0 captures what you're doing now. Layer 1 tracks the project. Layer 2 remembers *you* across every project — your tool preferences, coding style, patterns you always use.

Branch-Aware Memory

Switch branches, switch context. Automatically.

Working on feature/auth? Claude loads the auth overlay. Switch to feature/payments? Different context loads instantly. Merge a branch? Claude folds the overlay back into base memory…