What is Cortex?

Cortex is a local-first knowledge store designed for the way modern AI tooling actually works: long-running CLI agents, sessions that compact and lose context, work that spans many projects, and a constant need to remember things that the model itself cannot.

It is not a note-taking app. It does not try to replace any of the tools you already use to write. It sits underneath them, capturing the substrate of your work and exposing it back to AI clients through a stable, structured surface.

What it is

An inbox that any tool can append events to.
A normalizer that turns those events into atomic, addressable records.
A filesystem layout organised PARA-style: projects, areas, resources, archive.
A SQLite sidecar for full-text search, dedup, activation tracking, and indexes.
An MCP server exposing five canonical operations to AI clients.
A set of lifecycle hooks that bind it into session start, pre-compact, and stop events.
A distillation loop that turns accumulated records into higher-level summaries over time.

What it is not

It is not a hosted service. There is no server to log into, no account to create.
It is not a vector database. It uses SQLite FTS for retrieval; activation is tracked separately, not via embeddings.
It is not a wiki. There is no UI for browsing — the access path is AI clients calling MCP, or the filesystem directly.
It is not opinionated about which AI tools you use. Any client that speaks MCP can drive it.

The animating idea

Sessions with AI agents are stateful, but the state is fragile. Context windows compact. Sessions end. Tools restart. Most of what was useful in the session is lost the moment the next one begins.

Cortex assumes that the session is the wrong unit of memory. The right unit is the durable, addressable record — small, self-contained, and reusable across sessions, tools, and projects.

The job of the session is to produce events. The job of Cortex is to turn those events into records, and to put the right records back in front of the next session.