01
Progressive recall
3-layer retrieval (facts → summaries → full content) stops at the cheapest layer that answers — facts cost 5–15 tokens.
For Claude Code
Cortex
Memory that remembers what matters.
Cortex gives Claude Code persistent memory through a 3-layer retrieval system that returns only what's relevant, using ~66% fewer tokens on average. It automatically extracts structured facts from your memories and code into a queryable knowledge graph, with entity normalization and multi-hop traversal. Everything stays local — SQLite + Python, zero API calls, zero cost. It's the memory the rest of the family plugs into, and the project that named the lab.
$ pip install cortex-claude && cortex-claude setup What it does
6 things, done well 02
Knowledge graph
Auto-extracted structured facts with entity normalization, deduplication, and multi-hop traversal across connections.
03
Code graph
Tree-sitter indexes symbols across 7 languages — ~100 tokens to look one up vs 1000+ to read the file.
04
Smart auto-capture
Hooks capture results from your tools and a background daemon saves them in about 0.3 seconds.
05
Self-organizing clusters
Memories group into semantic sub-graphs with human-readable labels — browse topics top-down before drilling in.
06
Decay & temporal awareness
Unused memories fade, contradicted facts lose confidence, and timestamps keep recall fresh instead of stale.
What you need
- Works with
- Claude Code, via MCP
- Runs on
- macOS, Linux, Windows · needs Python 3.11+
- Price
- Free — zero API calls, zero cost
- Dashboard
- A web UI at localhost:37800
- Your data
- Stays local, in a file on your machine
