Claude gets smarter about you with every conversation.
Most AI conversations are disposable. Cortex makes them cumulative. It runs silently in the background — observing your preferences, capturing knowledge, learning from corrections — so every conversation builds on the last.
No commands required. Just use Claude normally. It learns.
Works on Cowork (Claude Desktop) and Claude Code.
v3 required you to say /remember. v4 doesn't wait.
| Behavior | What it does | Platform |
|---|---|---|
| Auto-recall | Loads your profile and project context at conversation start | Both |
| Passive observation | Silently learns preferences, corrections, and domain knowledge | Both |
| Contextual recall | Surfaces relevant knowledge when you mention a project/topic mid-conversation | Both |
| Auto-commit | Saves knowledge and observations when the conversation ends | Both |
| User profile | Persistent model of who you are and how you like to work | Both |
| Per-project config | Control capture aggressiveness per project via .cortex.json |
Both |
| Claude Code support | Drop-in CLAUDE.md instructions + optional hooks | Claude Code |
Conversation starts
→ Auto-recall loads your profile + project context
→ Claude adapts to your known preferences
Conversation happens
→ Passive observation accumulates new signals
→ Contextual recall surfaces relevant knowledge on mention
→ Claude adapts in real-time
Conversation ends
→ Auto-commit saves decisions, knowledge, and observations
→ Your profile and project nodes get smarter
Next conversation starts
→ Claude knows more. Repeats less. Helps better.
- Download the plugin zip
- Claude Desktop → Cowork tab → Customize → Upload custom plugin
- Select
cortex.zip - Start talking. Memory activates automatically.
Minimal setup (no hooks):
- Copy the contents of
claude-code/INSTRUCTIONS.mdinto your~/.claude/CLAUDE.md(global) or project-levelCLAUDE.md mkdir -p ~/Documents/Claude/memory- Start a new Claude Code session. Done.
With hooks (recommended):
- Do the minimal setup above
- Add the hooks from
claude-code/hooks.jsonto your~/.claude/settings.json - Hooks make auto-recall more reliable by feeding data at session init
At conversation start, Claude reads your user profile and checks for attention items. If you have overdue P0s or stale threads, it tells you. If you mention a known project, it loads context automatically. All silently — no "loading memory..." preamble.
During every conversation, Claude watches for:
- Corrections ("don't do that", "I prefer this") — highest priority, always saved
- Preferences (communication style, tool choices, working patterns)
- Domain knowledge (how your company works, industry terms, constraints)
- Relationship context (who's who, who owns what)
It never interrupts to capture these. It adapts in real-time and saves at conversation end.
When you mention a project, person, or topic with memory, Claude surfaces relevant knowledge naturally — "Heads up, Acme's procurement requires 3 vendor quotes" — without dumping a formal recall block.
When the conversation ends, Claude silently commits what was learned. No confirmation prompt. If the conversation was trivial, it skips. Corrections and preferences are always saved.
New in v4: a user node that accumulates knowledge about you:
- Communication preferences (terse vs. detailed, options vs. decisions)
- Working style (batching, cadence, delegation patterns)
- Corrections (things Claude should never do again)
- Domain expertise (what you know, what needs explaining)
- Relationships (who you work with, their roles)
- Tool preferences (which platforms you use)
This is what makes Claude feel like it "knows" you. It carries across all projects and both platforms.
All commands from v3 still work. v4 added subagent invocation; v4.2 adds shared-config commands and closing rituals.
| Command | What it does |
|---|---|
/remember |
Full session commit with confirmation |
/recall [project?] |
Dashboard (no arg) or specific project context |
/learn [node] [type?] [content] |
Quick knowledge capture |
/note [node] [content] |
One-liner fact |
/search [query] |
Cross-project search (delegates to memory-librarian for broad queries) |
/review |
Weekly synthesis digest |
/timeline [project?] |
Chronological activity |
/forget [node] |
Archive a project |
/cleanup |
Memory health audit |
These write to canonical files at ~/Documents/Claude/ that every plugin in the BrightWayAI marketplace reads. Capture once, all plugins benefit.
| Command | What it does |
|---|---|
/setup-identity |
Captures name, company, role, primary tools, communication defaults to ~/Documents/Claude/identity.md. Other plugins skip identity questions in their setups. |
/setup-voice |
Captures voice descriptors, banned phrases, sentence rhythm, hook patterns, sign-off style to ~/Documents/Claude/voice.md. Drafting plugins read from here. |
| Command | What it does |
|---|---|
/end-day |
5-min daily close — recap today, prompt for reflection, commit learnings to memory, optionally pre-stage tomorrow via plan-tomorrow. |
/end-week |
15-min Friday close — runs transcript-reviewer for uncaptured commitments, /cleanup for memory hygiene, /review for synthesis, prompts for weekly reflection, optionally pre-stages Monday's outreach via weekly-outreach. |
| Skill | Behavior |
|---|---|
observe |
Runs silently in every conversation. No trigger needed. Captures preferences, corrections, and domain knowledge. |
Commands also fire from natural language:
| Trigger | Skill |
|---|---|
| "save this", wrapping up, farewell | remember (auto-commit) |
| Conversation start, greeting, "catch me up" | recall (auto-recall) |
| "TIL", "gotcha:", "the trick is...", "I was wrong about" | learn |
| "note that", "jot down", "quick note" | note |
| "any gotchas with", "what's blocked", "how does X work" | search |
| "weekly review", "summarize my week" | review |
| "we're done with X", "archive X" | forget |
| "what have I been working on" | timeline |
| "clean up memory", "what's stale" | cleanup |
| "set up my identity", "configure my profile across plugins" | setup-identity |
| "set up my voice", "update my writing voice" | setup-voice |
| "wrapping up", "calling it a day", "end of day" | end-day |
| "Friday wrap-up", "close out the week", "end of week" | end-week |
Cortex ships two specialist subagents that handle heavy memory work off the main conversation thread. Other plugins (or Cortex's own commands) can delegate to them via the Task tool.
| Subagent | Purpose | Used by |
|---|---|---|
memory-librarian |
Search, synthesize, and deduplicate across all working-memory files for broad / cross-cutting queries. Returns a structured Summary / Source Entries / Open Threads / Confidence response. Read-only. | /search (cortex) for broad queries; any other skill that needs cross-node memory synthesis |
transcript-reviewer |
Read recent Granola call transcripts and surface commitments the user made that aren't already tracked in CRM tasks or Cortex memory. Returns a delta — only what's missing from existing tracking. | Manual / scheduled weekly run |
Subagents inherit parent tools at runtime, so they work with whichever connectors (Granola, CRM, etc.) the user has connected. Tool allowlists are kept tight where possible (e.g., memory-librarian uses only Read/Grep/Glob).
To invoke a subagent directly via the Task tool (Cowork or Claude Code):
Use the Task tool with subagent_type="memory-librarian"
and pass: query="[your question]"
The agents are registered automatically when Cortex is installed — they appear in Claude's subagent_type enum alongside built-in agents.
- Decisions made and reasoning
- Open threads with staleness tracking
- Blockers and who/what is blocking
- Artifacts created (docs, decks, proposals)
- Next actions (P0/P1/P2/WAITING)
| Type | Example |
|---|---|
| Insight | "Churn is an onboarding problem, not a product problem" |
| Lesson | "Same-day proposals fail. 48-hour tailored decks close 3x better" |
| Model | "Their approval: dept head → finance → VP → procurement" |
| Gotcha | "Fiscal year starts April, not January — adjust all budget timing" |
| Recipe | "For exec buy-in: lead with their metric, show the gap, propose one action" |
| Correction | "Not price-sensitive — they need ROI framing for internal approval" |
| Type | Example |
|---|---|
| Preference | "Prefers terse responses — told Claude to stop summarizing" |
| Correction | "Don't add emojis to professional communications" |
| Pattern | "Starts mornings with email triage, then deep work blocks" |
| Domain | "Deep expertise in AI go-to-market — skip 101 explanations" |
| Relationship | "Reports to [Name] (CEO), weekly 1:1s on Mondays" |
Control behavior per project via .cortex.json in the project root:
{
"node": "client:acme-corp",
"capture": "aggressive",
"auto_recall": true,
"auto_commit": true,
"observe": true
}| Capture Level | Behavior |
|---|---|
"aggressive" |
Capture everything. Use for high-value client work. |
"normal" |
Standard capture. Default. |
"minimal" |
Only explicit commands. No auto behaviors. |
See cortex.config.md for the full spec.
Memory lives on your computer at ~/Documents/Claude/memory/. Shared between Cowork and Claude Code.
~/Documents/Claude/memory/
├── DASHBOARD.md ← Master index
├── user.md ← Your profile (NEW in v4)
├── archive/ ← Archived nodes
├── client/
│ ├── acme-corp.md
│ └── northstar.md
├── bizdev/
│ └── partnerships.md
├── strategy/
│ └── q2-growth.md
├── hiring.md
└── brand.md
All files are plain markdown. Human-readable. Editable. Backupable.
Use kebab-case. Organize however fits your work:
| Pattern | Example |
|---|---|
| Client work | client:acme-corp |
| Business dev | bizdev:stripe-partnership |
| Internal ops | company-ops, hiring, finance |
| Strategy | strategy:q2-growth |
| Products | onboarding-program, crm-dashboard |
| Learning | learning:sales-ops |
| Domain | domain:healthcare-compliance |
| Infrastructure | infra:data-pipeline |
| Personal | personal |
| User profile | user (auto-managed) |
| Feature | Cowork | Claude Code |
|---|---|---|
| Plugin system | Native | Via CLAUDE.md |
| Auto-recall | Skill auto-fire | CLAUDE.md instructions + hooks |
| Auto-commit | Skill auto-fire | CLAUDE.md instructions |
| Passive observation | Always on | Always on |
| Contextual recall | Always on | Always on |
| User profile | Shared | Shared |
| Per-project config | Supported | Supported |
| Explicit commands | All 10 | All 10 |
| Memory files | Shared location | Shared location |
Both platforms read/write the same memory files. Learn something in Cowork → Claude Code knows it. And vice versa.
- Passive observation engine: Claude silently learns about you during every conversation
- User profile node: Persistent model of preferences, corrections, patterns, domain expertise
- Auto-recall: Context loads at conversation start without commands
- Auto-commit: Knowledge saves at conversation end without commands
- Contextual recall: Relevant knowledge surfaces mid-conversation on mention
- Claude Code support: Full integration via CLAUDE.md instructions and optional hooks
- Per-project config:
.cortex.jsonfor per-directory behavior control - Silent mode: Auto-triggered commits produce no output (unless creating new nodes)
- The learning loop: Every conversation makes the next one better
- File-based storage at
~/Documents/Claude/memory/ - Two-tier structure: DASHBOARD.md + individual node files
- Dynamic directory creation from node prefixes
- Archive support
- Business-operator friendly language and examples
- Strategy/planning node types
- Knowledge-first redesign with six knowledge types
- Added
/learnand/notefor quick capture /reviewweekly digest
- Added
/search,/forget,/timeline,/cleanup - Priority system, staleness tracking, cross-project signals
- Initial release with
/rememberand/recall