pyramidex

A CLAUDE.md that loads on demand.

Pyramidex stores your AI configuration in a knowledge graph. Each Claude Code session starts with a small index, and the tool fetches only the rules, memories, and workflows the current task actually needs. Less context per session, no duplicate rules drifting between files, a config that keeps up as it grows.

Built for Claude Code only - other AI hosts are out of scope

Why use Pyramidex?

For the data model and design rationale, see docs/knowledge-design-guide.md

What you get:

1. Lazy, tiered loading. Sessions start with a ~275-token domain index; only the domains the task touches are fetched
2. No duplication. A rule that applies to two contexts is one node with two edges, not two copies that drift apart
3. Self-describing. The graph carries its own schema and bootstrap query, no external docs to keep in sync
4. Atomic edits. Adding or changing a rule is one write, not "edit the file, update the index, fix the description"
5. Portable when hosted in the cloud. Same rules and memories on every machine, no dotfile sync

What it costs:

1. A graph database. Managed (Aura free tier works) or self-hosted, either way another moving part
2. An MCP server per machine. One-time install, but credentials must be configured wherever you use it
3. A learning curve. You edit config by talking to Claude, not by editing a file

When it is worth it:

Your CLAUDE.md feels crowded, you have rules duplicated across contexts, or you work from more than one machine. If your config is a dozen lines in a single file on one laptop, stick with plain Markdown

Getting started

git clone https://github.com/morooshka/pyramidex
cd pyramidex
claude

Then ask Claude:

Follow docs/prompts/setup.md and set up Pyramidex for me

First-time setup runs about 10-15 minutes, mostly waiting for the Neo4j Aura instance to spin up. Claude handles the rest

License

MIT - see LICENSE