Frequently Asked Questions

Cortex is the memory and control layer for Claude. It runs on your Mac, captures the substantive moments from your sessions, and keeps your memories, agents, skills, and projects in sync across Claude Code, Claude Desktop, and Cowork. The goal: Claude stops starting from scratch every time — and what it knows is one thing, not scattered per surface.

Anyone who uses Claude daily — developers, designers, writers, researchers, therapists, lawyers, accountants, teachers, marketers. If you've ever re-explained the same preference to Claude more than twice, or wished what Claude knew in one place carried to another, Cortex is for you.

Claude has memory now — but it's per surface, and you maintain it by hand. Claude Code keeps one memory, Claude Desktop another, your phone a third, and none of them follow you between machines.

Cortex is the layer underneath that makes it one memory, kept in sync everywhere Claude runs. It's also a real memory system: it captures automatically, lets unused memories decay into a restorable cold archive, and consolidates what recurs into durable rules — the difference between a pile of notes and something Claude actually knows.

macOS 14 (Sonoma) and newer, on both Apple silicon and Intel. Cortex is a downloadable public beta now, and works with Claude Code, Claude Desktop, and Cowork. Windows and Linux aren't on a committed timeline yet.

The whole local control panel is free and downloadable now — memories, agents, skills, projects, automatic capture, decay, and consolidation, all on your Mac.

A hosted Cloud Sync tier (zero-setup sync to claude.ai and mobile, about $49/year) is coming. Power users can also self-host the Connector free today on their own Cloudflare. See the pricing section for the full breakdown.

Rob Stout, an independent designer and developer based in the US. Cortex is an indie product — no VC, no quarterly OKRs pushing the roadmap away from the people using it.

Likely, but not yet. When we have enough active users for a community to be useful instead of empty, we'll open one up.

When a Claude Code session ends, Cortex saves the substantive moments — decisions, corrections, preferences — as short-lived "episodic" memories, so you don't have to say "remember this." It's capped and de-duplicated. To stop it, set capture to Off in Settings → AI. Anything captured goes to trash for 7 days before removal, and memories that fade settle into a cold archive you can restore from at any time. Nothing is deleted out from under you.

No. Cortex captures and files memories on its own and surfaces what's relevant each session. You can still open the app to browse, pin, edit, or delete — and pinned/semantic memories mirror to readable ~/.cortex/memory/*.md files you can edit in any editor — but day to day it runs itself.

Cortex organizes memories into six tiers, modeled on how brains actually handle memory:

• Sensory — raw stream of the current session, mostly discarded
• Working — what Claude is actively thinking with right now
• Episodic — specific events, corrections, and decisions
• Semantic — generalized rules and facts that cross every project
• Procedural — skills and workflows (the how)
• Long-term — everything consolidated for the long haul

Memories move between layers over time: specific events get promoted to general rules; rules that aren't reused decay. It's automatic.

When you start a session, Cortex sees the working directory, open files, and conversation context. It ranks memories by relevance to that context — project scope, recency, reuse frequency, and user-pinned importance — and surfaces the top matches to Claude. You can always promote or pin a memory manually if Cortex misses one.

Cortex provides always-on memory through MCP — the Model Context Protocol, an open standard from Anthropic that lets Claude talk to local tools and data sources. Cortex registers SessionStart and SessionEnd hooks, so Claude Code reads your memory at the start of every session and Cortex captures the substance at the end.

You can confirm everything's wired up with Verify Connection in Settings. Installing Cortex sets up the MCP connection for both Claude Desktop and Claude Code automatically — you don't configure anything by hand.

Everything lives on your Mac: a SQLite database at ~/.cortex/cortex.db, plus a readable markdown mirror at ~/.cortex/memory/*.md. Active memory is what Cortex surfaces to Claude. Memories you stop using cool into a cold archive (kept, just out of the way). Deleted items sit in trash for 7 days. You can clear by age or by project, and restore from archive or trash, in Settings → Storage.

Because the same agent, skill, project, or memory was edited in two places at once — say, in the Cortex app and directly in ~/.claude/…. Nothing was overwritten. Cortex shows a conflict sheet that leads with a diff of what actually changed and lets you Keep Cortex's version, Keep the file's version, or Keep both. Sync is additive-only: it can flag and pause, but it can't destroy data.

Yes. Cortex gives you a proper export of every memory with its metadata and timestamps — portable and re-importable. Your pinned and semantic memories also live as readable markdown at ~/.cortex/memory/*.md, which you can open or back up in any editor. It's a real export, not a copy-paste chore.

Cortex snapshots your memory database on a schedule so you can roll back. To restore, open Settings → Backups, pick a snapshot, and confirm. Session transcripts are retained per your retention setting; reduce or clear them there too.

Cortex's core — capturing, organizing, and surfacing your memories — runs entirely on your Mac and is free. It uses no AI to do that.

A few optional features can call a language model for a nicer result: Explain with Claude (plain-language summaries of a sync conflict) and the Enhanced capture level (distills cleaner memories at the end of a session). When you turn one on, Cortex runs it through your own Claude subscription using the Claude Code you already have installed. Cortex has no API key and never charges you for AI. If you don't have Claude Code, these features simply don't appear — everything else works exactly the same.

One thing to know: from June 15, 2026, Anthropic meters programmatic (CLI) calls against a separate monthly credit included with your plan (about $20 on Pro, $100–200 on Max). You claim it once in your Claude account settings. If the AI features ever stop responding, that credit is the first thing to check.

In Settings → AI you choose how memories are made. Off: no automatic capture. Standard (default): on-device heuristics decide what's worth keeping — free, private, no AI. Enhanced: once per session, Cortex asks Claude (through your own subscription) to distill cleaner memories — capped at ~10 sessions/day, a few cents of your plan's credit each.

No. Cortex is a single Mac app. It doesn't install other software. Its optional AI features work with Claude Code if you already have it; if you don't, they're just hidden. Nothing to set up, nothing bundled.

Today Cortex is built for Claude only — it connects through MCP (Model Context Protocol), which Anthropic supports natively in Claude Desktop and Claude Code. MCP is an open standard, so other assistants could integrate with Cortex in the future. Nothing is planned yet.

No — not unless you choose it. By default Cortex makes no network calls and has no servers. Two opt-in exceptions: the optional AI features (which run through your own Claude Code, to Anthropic, on your subscription — Cortex never sees that data), and the self-host Connector, where you run a small worker on your own Cloudflare account to reach your memory from claude.ai and mobile. You can also export your memories yourself. Beyond that, nothing leaves.

Yes. Forget everything in Settings clears your local memory in one step; a blast-radius guard makes you confirm before anything irreversible. You can also export first, or delete ~/.cortex/ in Finder. (See the full privacy policy.)

Cortex reads the conversation context that Claude shares with it through MCP — specifically, the corrections and rules you ask Cortex to remember. It does not log, transmit, or analyze your broader conversation content. Your chats with Claude stay between you and Claude.

No. Never. Cortex exists to be your memory — not a training set. We do not share, sell, or license your memories to anyone, for any purpose, including AI training.

The first time you turn on always-on memory, macOS shows a prompt that says "Cortex would like to access data from other apps." Click Allow.

That prompt is macOS asking permission for Cortex to write a single line to ~/.claude/settings.json — the file Claude Code reads at startup. That one line is what lets Claude Code call Cortex at the start of every session, so the context Cortex has learned carries across sessions automatically.

Cortex only touches the entry it added. Your other Claude Code settings stay untouched. You can revoke the permission anytime in System Settings → Privacy & Security → App Management, or remove the hook itself from Cortex's settings.

If you click Don't Allow, Cortex still works — you'll just need to manually share context with Claude when you want it. Always-on memory is the recommended setup.

No. Your waitlist email is used only for the updates you signed up for. No newsletters, no partners, no third-party sharing. See the privacy policy.

Cortex updates itself automatically (via Sparkle) and tells you when a new build is ready. You can also check manually in Settings → About → Check for Updates, where release notes live.

Email [email protected] — every message is read by a human. Please include your version and build (Settings → About) so we can reproduce what you're seeing. You can also use in-app feedback from the menu bar.