Your AI agent remembers
Cortex gives Claude persistent memory and structured reasoning that compounds across sessions. Decisions, patterns, and context — automatically captured, instantly searchable.
0
Re-explanation
Prior decisions loaded at session start
4
Access paths
MCP Server, Edge Function, Import, CLI
24+
MCP tools
Search, create, traverse, capture
Auto
Capture
Hooks save notable changes for you
Every session starts from scratch
No memory between sessions
Your AI agent forgets everything when the session ends. That architecture decision you explained in detail? Gone. The gotcha that took 20 minutes to debug? Gone. Every session is a blank slate.
You repeat yourself constantly
"We use RLS for tenant isolation." "The auth chain goes through proxy.ts." "Don't use CREATE OR REPLACE on shared functions." The same context, re-explained every session.
Reasoning chains break
Complex analysis that spans multiple sessions — migration planning, architecture refactors, incident investigations — loses all intermediate reasoning when the session ends.
A brain that grows with every session
Cortex is the brain of the Composure organism. It captures decisions, patterns, and reasoning automatically via PostToolUse hooks — no commands needed. A knowledge graph of memory nodes and edges grows with every session, linking reasoning to projects and tasks via an entity registry. Sequential thinking chains persist across sessions, letting you pick up complex analysis exactly where you left off. And memory isn't just recall — it's shared state. Blueprints, one task queue, campaign locks, and memory itself live on disk where every session can read them, so two terminals on the same machine coordinate instead of colliding. Free tier runs locally on SQLite. Pro tier syncs across devices with semantic search.
See it in the field. Two sessions worked the same product on the same day: one read the other's in-progress blueprint and stayed out of its lane; when a misfired heartbeat told the wrong session to take over a live campaign, it declined. The store behind that coordination holds 12,927 memory nodes and 88 persisted thinking sessions. Both excerpts are public:
$# Session starts — Cortex auto-loads context$[cortex] Loaded 12 memory nodes for composure-web$[cortex] 3 active thinking sessions found$# You work normally — Cortex captures automatically$[cortex] Captured: migration safety pattern (from file change)$[cortex] Captured: RLS policy decision (from sequential thinking)$# Next session — everything is there$search_memory({ query: "RLS tenant isolation" })$→ 4 nodes: policy pattern, migration decision, test approach, gotcha$# Pick up a reasoning chain from last week$get_thinking_session({ id: "arch-refactor-2026-04" })$→ 8 thoughts, last updated 3 days ago$→ Resuming from thought #8...How it works
Auto-capture via hooks
SessionStart loads project and global memories. PostToolUse captures notable file changes, Context7 findings, and architecture patterns as memory nodes — no manual commands needed.
Knowledge graph grows
Memory nodes and edges form a project-aware graph. Decisions link to the files they affected. Patterns link to the projects that use them. Everything is connected and traversable.
Sequential thinking persists
Chain-of-thought reasoning with full context — thought content, tool inputs, conversation snapshots. Resume complex analysis across sessions, even weeks apart.
Four doors, one brain
Access Cortex through MCP Server (Claude Code), Edge Function (web), Direct Import (scripts), or CLI (terminal). Same memory, any access path.