🧠 Brain 2.0 | OpenClaw Plugin

> Your second brain. Capture anything, organize nothing.

> 💛 Like Brain? Tips help keep development going: D6i8TeepmrGztENxdME84d2x5UVjLWncat (DOGE)

---

Table of Contents

• What is Brain 2.0
• Quick Start
• Everyday Use
• Your Buckets
• Staying on Track
• Managing Your Brain
• Going Deeper
• Under the Hood
• Reference
• Troubleshooting
• Roadmap

---

What is Brain 2.0

Brain 2.0 is a behavioral support system that extends your biological memory. The philosophy is simple: capture everything, organize nothing. You drop raw thoughts (a name, an idea, a reminder, a to-do) and Brain automatically classifies them, files them into the right bucket, detects duplicates, and even creates persistent reminders for time-sensitive items. You never have to decide where something goes. Just drop it and move on.

Brain also keeps you on track with bite-sized digests throughout the day: morning briefs, midday checks, afternoon wraps, and night wind-downs. Each digest is intentionally small, Pinky 🤙, so you can scan it and act in seconds, not minutes. There's also a weekly review to zoom out and see the bigger picture.

Brain offers two ways to interact. Slash commands (/brain drop, /brain search, etc.) execute instantly without touching the AI: they're faster, cheaper, and critically, don't add to your chat context. Your conversation window stays clean for the work that matters. Just type /drop Pick up prescription tomorrow and it's captured, classified, and reminded, all at zero AI token cost.

For richer interactions, agent tools (brain_drop, brain_search, etc.) work through natural conversation, letting the AI reason about your query and provide contextual follow-ups. For example: "Did I already drop something about picking up my prescription?" and the AI searches your brain, finds the match, and confirms the context back to you.

---

Quick Start

1. Install

Brain 2.0 is a separate plugin and does not come bundled with OpenClaw.

Option A: From GitHub (recommended)

git clone https://github.com/quackstro/openclaw-brain.git ~/.openclaw/extensions/brain
cd ~/.openclaw/extensions/brain
npm install

Option B: From npm (coming soon)

openclaw plugins install @quackstro/openclaw-brain

Then restart OpenClaw to load the plugin:

openclaw gateway restart

Verify it loaded:

openclaw plugins list
# Should show: brain - Brain 2.0

2. Configure

Add your embedding API key to the Brain config in your OpenClaw configuration (~/.openclaw/openclaw.json):

{
  "plugins": {
    "brain": {
      "embedding": {
        "apiKey": "your-api-key-here"
      },
      "classifier": {
        "apiKey": "your-gateway-token"
      }
    }
  }
}

3. First Drop

Slash command (fastest, no AI, no context cost):

/drop Call dentist next Tuesday

Agent tool (programmatic, used by the AI when you interact via chat):

> brain_drop: "Call dentist next Tuesday"

Natural language (conversational, the AI decides to use Brain for you):

> "Remember that I need to call the dentist next Tuesday"

All three do the same thing. Brain captures it, classifies it as admin (appointment), and creates a reminder for Tuesday. ✅

---

Everyday Use

Slash Commands

Brain registers a /brain slash command that works directly in Telegram (and other channels) without invoking the AI agent.

> 💡 Why slash commands matter: > > - ⚡ Instant: No LLM round-trip. Results come back in milliseconds, not seconds. > - 💰 Zero token cost: Slash commands don't consume any AI tokens. > - 🧹 No context pollution: The biggest advantage. Slash commands don't add messages to your chat history. When you drop a thought via /brain drop, it doesn't eat into your conversation window. Your chat context stays focused on the current task, not cluttered with brain operations. > - 🔒 Deterministic: Same input, same output. No LLM interpretation or hallucination risk.

Use slash commands for quick captures, lookups, and status checks. Use the agent tools (via natural chat) when you need conversational follow-ups or complex queries like "search my brain for that thing my neighbor mentioned about 3D printing."

/brain

Show the Brain dashboard with bucket counts and DND status.

🧠 Brain 2.0 Dashboard

  people: 2
  projects: 37
  ideas: 11
  ...

🔔 DND: OFF

Commands: drop, search, stats, dnd

/drop <text>

The fastest way to capture a thought. Shortcut for /brain drop.

/drop Call dentist about appointment next Tuesday

→ ✅ Captured [ToDo] (with ID)

/brain drop <text>

Same as /drop, for when you want the full command.

/brain drop Pick up prescription tomorrow

→ ✅ Captured (with ID)

/brain search <query>

Semantic search across all buckets. Returns top 5 matches with scores.

/brain search mobile app project status

→

🔍 Found 5 results:

1. [projects] Mobile App Redesign (92%)
2. [projects] Payment Integration (78%)
3. [admin] Weekly report schedule (71%)
...

/brain stats

Show record counts per bucket (same as the dashboard stats section).

/brain dnd [on|off|status]

Control Do Not Disturb mode directly.

/brain dnd on      → 🔇 Do Not Disturb enabled.
/brain dnd off     → 🔔 Do Not Disturb disabled.
/brain dnd status  → 🔇 DND is ON: Manual override
/brain dnd         → (same as status)

> 💡 Tip: For routine Brain operations (quick drops, status checks, DND toggles), always prefer slash commands. Reserve agent tools for when you need the AI to reason about your query, e.g., "what was that project idea I had last week about security cameras?"

Natural Language

You don't need slash commands or tool names. Just talk to your assistant:

"Met Sarah from Acme Corp at the conference, she's working on ML infrastructure"
"Remind me to renew my passport before March"
"What if we built a CLI tool that generates changelogs from git history?"
"Gym appointment Thursday 3pm"

The AI recognizes these as Brain-worthy and uses the right tool automatically.

Bracket Tags

You can prefix your drops with bracket tags to explicitly tell Brain what kind of action this is. Tags are case-insensitive and are stripped from the text before classification.

| Tag | Intent | What happens | |-----|--------|-------------| | [ToDo] | todo | Tagged as a task, logged to audit | | [Reminder] | reminder | Creates a persistent cron reminder with Telegram buttons | | [Buy] | purchase | Tagged as a purchase item with amount extraction | | [Call] | call | Tagged as a call follow-up with contact extraction | | [Book] | booking | Creates an appointment reminder (like Reminder) |

Examples:

[ToDo] buy groceries after work
[Reminder] call mom at 5pm tomorrow
[Buy] new headphones, budget $200
[Call] follow up with Dr. Smith about lab results
[Book] dinner reservation Friday 7pm at Olive Garden

If no bracket tag is provided, Brain uses the classifier's detected intent and keyword heuristics to infer the action automatically.

---

Your Buckets

Brain organizes your thoughts into 9 buckets:

| Bucket | What goes here | Example | |--------|---------------|---------| | 🧑 people | Contacts, relationships, follow-ups | "Met Sarah from Acme, follow up next week" | | 📁 projects | Active work, tasks, deadlines | "Website redesign, waiting on design mockups" | | 💡 ideas | New concepts, what-ifs, opportunities | "What if we built a personal CRM?" | | 📋 admin | Appointments, errands, logistics | "Dentist appointment Thursday at 2pm" | | 📄 documents | Articles, references, resources | "Great article on distributed systems by Martin Kleppmann" | | 🎯 goals | Long-term objectives, milestones | "Run a half marathon by September" | | 🏥 health | Medical, fitness, nutrition, wellness | "Blood pressure was 120/80 at last checkup" | | 💰 finance | Bills, investments, expenses, budgets | "Electric bill due on the 15th, $142" | | 💬 conversations | Auto-captured facts, decisions, preferences, todos | Distilled from chat sessions by the auto-capture pipeline |

Each bucket has its own schema with fields tailored to its purpose (e.g., people records track company, contactInfo, and lastInteraction; finance records track amount, currency, and recurring schedules). The conversations bucket has structured columns (type, entities, temporal, confidence, source_session) enabling hybrid SQL + vector search.

There are also 3 system tables: inbox (pending drops), needs_review (low-confidence items), and audit_trail (full history).

---

Staying on Track

Digests (Pinky 🤙)

> Digests are intentionally small. Bite-sized summaries you can scan and act on in seconds, not minutes. This is a core design principle: if a digest requires effort to read, it's failed. Word limits are enforced per digest type to keep them tight and actionable.

Brain generates 5 types of digests to keep you on track throughout the day:

☀️ Morning Brief (default: 150 words)

• Top 3 actions for today (urgent + overdue items)
• Overdue summary
• Items waiting for your review (>24h old)

🕐 Midday Check (default: 100 words)

• Items due today that aren't started
• Stuck items (no update in 3+ days)

🌆 Afternoon Wrap (default: 120 words)

• What moved today (recent activity)
• What carries to tomorrow

🌙 Night Wind-down (default: 100 words)

• Tomorrow's priorities
• Pending follow-ups (especially people)

📊 Weekly Review (default: 250 words)

• Bucket health overview (overdue, stuck, active counts)
• This week's focus (top 3 actions)
• Stuck items that need unblocking
• Wins from the past week

Daily digests are template-based (no LLM cost). They pull data directly from your buckets and format it using predefined templates.

Do Not Disturb

Brain's DND system controls when digests and notifications are sent. All digests respect DND mode: if DND is active, the digest is recorded as skipped and can be caught up later.

Auto-Quiet Hours: By default, Brain is silent from 10:00 PM to 7:00 AM (Eastern Time). During quiet hours, digests are skipped and recorded. You can customize the window in config.

Manual Override:

• brain_dnd(action: "on"): Immediately silences all Brain notifications. Overrides auto-quiet settings.
• brain_dnd(action: "off"): Resumes notifications. If digests were skipped during DND, Brain tells you what you missed.

Recovery: When you turn DND off, Brain reports which digests were skipped so you can generate a catch-up summary if needed. It does not flood you with all the missed digests; just a single summary.

DND state is persisted to ~/.openclaw/brain/dnd-state.json and survives restarts.

---

Managing Your Brain

Searching

Use brain_search to find anything across your Brain by semantic similarity.

brain_search("Sarah from Acme")
→ Found 2 results:
  1. [people] Sarah, Acme Corp ML Infrastructure (87%)
  2. [projects] Acme Partnership (62%)

You can also limit to a specific bucket:

brain_search(query: "headphones", bucket: "finance")

Fixing and Moving Items

Use brain_fix to inspect, move, trash, update, or merge items.

Show details (omit the correction):

brain_fix(id: "abc123")
→ 📋 Sarah, Acme Corp
    Bucket: people
    ID: abc12345-...
    context: Met at conference, works on ML infrastructure
    nextActions: ["Send follow-up email"]
    ...

Move to a different bucket:

brain_fix(id: "abc123", correction: "→ projects")
→ ✅ Moved "Sarah" from people → projects

Delete permanently:

brain_fix(id: "abc123", correction: "→ trash")
→ 🗑️ Deleted "Sarah" from people.

Update the next action:

brain_fix(id: "abc123", correction: 'action: "Schedule coffee chat"')
→ ✅ Updated next action for "Sarah": Schedule coffee chat

Merge two items:

brain_fix(id: "abc123", correction: "merge def456")
→ ✅ Merged "Other Sarah Note" into "Sarah". Combined 4 entries, 3 actions.

Viewing History

Use brain_audit to see what happened to any item:

brain_audit(inputId: "abc12345")
→ 📋 Audit Trail (3 entries):
  [2025-06-01T10:30:00] captured: Captured drop: "Call dentist Tuesday"
  [2025-06-01T10:30:01] classified: Classified as admin (0.94): "Dentist Appointment"
  [2025-06-01T10:30:02] routed: Routed to admin with confidence 0.94

Triaging Your needs_review Queue

When items land in needs_review (because the classifier wasn't confident enough), here's a practical workflow:

1. Check stats: brain_stats to see how many items are pending review. 2. Search needs_review: brain_search(query: "...", bucket: "needs_review") or use openclaw brain list needs_review to see them all. 3. Inspect each item: brain_fix(id: "...") to see details and the classifier's best guess. 4. Route it: brain_fix(id: "...", correction: "→ projects") to send it to the right bucket. 5. Or trash it: brain_fix(id: "...", correction: "→ trash") if it's noise.

---

Going Deeper

Full Configuration Reference

All options live under the brain key in your OpenClaw plugin config.

Embedding

| Option | Type | Default | Description | |--------|------|---------|-------------| | embedding.apiKey | string | required | API key for generating embeddings. Supports OpenAI and Google Gemini keys (auto-detected from key prefix; keys starting with AI use Gemini). | | embedding.model | string | text-embedding-3-small | Embedding model. Options: text-embedding-3-small, text-embedding-3-large, gemini-embedding-001 | | embedding.baseURL | string | - | Custom base URL for OpenAI-compatible embedding APIs |

Classification

| Option | Type | Default | Description | |--------|------|---------|-------------| | classifier.apiKey | string | - | API key / gateway token for the classification LLM. If omitted, classification is disabled and drops stay in the inbox. | | classifier.model | string | claude-sonnet-4-20250514 | LLM model used for classifying thoughts | | classifier.confidenceThreshold | number | 0.80 | Minimum confidence (0.0-1.0) required to auto-route a drop to a bucket. Below this, the item goes to needs review. |

Storage

| Option | Type | Default | Description | |--------|------|---------|-------------| | storage.dbPath | string | ~/.openclaw/brain/lancedb | Path to the LanceDB database directory |

Digests

| Option | Type | Default | Description | |--------|------|---------|-------------| | digests.enabled | boolean | - | Master switch for the digest system | | digests.timezone | string | - | IANA timezone for digest scheduling (e.g., America/New_York) | | digests.schedule.morning.enabled | boolean | - | Enable the morning digest | | digests.schedule.morning.time | string | - | Time to send (e.g., "07:30") | | digests.schedule.morning.maxWords | number | 150 | Soft word limit for morning digest | | digests.schedule.midday.enabled | boolean | - | Enable midday check-in | | digests.schedule.midday.time | string | - | Time to send | | digests.schedule.midday.maxWords | number | 100 | Soft word limit | | digests.schedule.afternoon.enabled | boolean | - | Enable afternoon wrap-up | | digests.schedule.afternoon.time | string | - | Time to send | | digests.schedule.afternoon.maxWords | number | 120 | Soft word limit | | digests.schedule.night.enabled | boolean | - | Enable night wind-down | | digests.schedule.night.time | string | - | Time to send | | digests.schedule.night.maxWords | number | 100 | Soft word limit | | digests.schedule.weekly.enabled | boolean | - | Enable weekly review | | digests.schedule.weekly.day | string | - | Day of week (e.g., "sunday") | | digests.schedule.weekly.time | string | - | Time to send | | digests.schedule.weekly.maxWords | number | 250 | Soft word limit |

Auto-Capture

| Option | Type | Default | Description | |--------|------|---------|-------------| | autoCapture.enabled | boolean | false | Master switch for conversation auto-capture | | autoCapture.level | string | "standard" | Aggressiveness tier: off, facts, standard, full | | autoCapture.extractionModel | string | "gemini-2.0-flash" | LLM model for turn extraction | | autoCapture.flushAfterTurns | number | 10 | Flush after this many user turns | | autoCapture.idleFlushMs | number | 300000 | Flush after idle period (ms). Default 5 minutes | | autoCapture.maxBufferTurns | number | 20 | Max turns in the ring buffer before eviction | | autoCapture.bucket | string | "conversations" | Target bucket for extracted items | | autoCapture.minConfidence | number | 0.6 | Skip extracted items below this confidence | | autoCapture.extractionApiKey | string | — | API key for extraction LLM (Gemini key or gateway token). Falls back to embedding.apiKey if omitted | | autoCapture.gatewayUrl | string | "http://127.0.0.1:18789" | Gateway URL for gateway-based extraction models |

Search

| Option | Type | Default | Description | |--------|------|---------|-------------| | search.excludeBuckets | string[] | [] | Buckets to exclude from default brain_search. Add "conversations" to silence auto-capture noise |

Do Not Disturb

| Option | Type | Default | Description | |--------|------|---------|-------------| | dnd.autoQuiet.enabled | boolean | true | Automatically silence digests during quiet hours | | dnd.autoQuiet.from | string | "22:00" | Start of quiet hours (24h format) | | dnd.autoQuiet.to | string | "07:00" | End of quiet hours (24h format) | | dnd.timezone | string | "America/New_York" | IANA timezone for quiet hour calculations |

Usage Awareness

| Option | Type | Default | Description | |--------|------|---------|-------------| | usageAware.enabled | boolean | - | Enable usage-aware backoff for API calls | | usageAware.backoffThreshold | number | - | Token usage threshold to trigger backoff |

Action Routing

After classification and bucket routing, Brain's action router inspects each drop for actionable intent. It handles:

| Intent | What happens | |--------|-------------| | reminder | Creates a persistent cron reminder with Telegram inline buttons (Dismiss / Snooze). Nags every 5 minutes until you act. | | booking | Same as reminder. Extracts the appointment time and sets up notifications. | | todo | Tags the item and logs it to the audit trail. | | purchase | Tags the item with amount info from extracted entities. | | call | Tags the item with contact info from extracted entities. | | payment | Resolves recipient from Brain's people bucket, creates approval UI with Approve/Edit/Dismiss buttons. Auto-executes for ≤1 DOGE to known recipients. |

Payment Actions

When Brain detects a payment intent (e.g., "send 5 DOGE to Alice"), the action router:

1. Resolves the recipient — Searches the people bucket for a matching contact with a DOGE address 2. Scores confidence — Based on recipient match quality, amount, and context 3. Routes by policy:

• Auto-execute (score ≥ 0.9, amount ≤ 1 DOGE, known recipient): Sends immediately via wallet_send
• Propose (score ≥ 0.5): Shows Telegram approval UI with Approve / Edit / Dismiss buttons
• Reject (score < 0.5 or unknown recipient): Notifies user of resolution failure

4. Race guard — Prevents duplicate sends from rapid button taps (status checked before and after wallet call) 5. Wallet lock handling — If wallet is locked, queues action as awaiting-unlock with retry on unlock event

Pending actions are stored as JSON in ~/.openclaw/brain/pending-actions/.

Time extraction uses an LLM call to parse natural language like "next Tuesday at 3pm" or "every Monday morning" into precise dates and cron expressions.

Intent priority: Explicit bracket tag (e.g., [Reminder]) > classifier-detected intent > keyword heuristics (regex patterns for times, days, and reminder words).

Deduplication

Before creating a new record, Brain checks for duplicates using cosine similarity on embedding vectors. If an existing record in the same bucket has a similarity score ≥ 0.92 (92%), the new drop is auto-merged into the existing record instead of creating a duplicate.

Merging appends the new information to the existing record's entry log and combines next-actions (with deduplication).

Auto-Routing

After classification, the router (the "Bouncer") decides what happens:

• Confidence ≥ threshold (default 0.80): The drop is automatically routed to the target bucket. A structured record is created with all the classified fields.
• Confidence < threshold or unknown bucket: The drop goes to the needs_review queue. You can review it later and manually route it with brain_fix.

Auto-Capture v2 (Conversation Distillation)

Instead of embedding every raw chat message (noisy and wasteful), Brain's auto-capture v2 pipeline buffers conversation turns and periodically distills them into high-quality structured items.

How it works:

1. Buffer: User and assistant turns accumulate in a per-session in-memory ring buffer. 2. Flush: When a trigger fires (idle timeout, turn count, or session end), the buffer is flushed. 3. Extract: A lightweight LLM (Gemini Flash or Haiku) distills the buffered turns into structured items: facts, decisions, preferences, and todos. 4. Embed & Store: Each extracted item gets its own embedding and is stored in the conversations bucket with typed metadata. 5. Skip noise: If the LLM finds nothing worth extracting (e.g., casual chit-chat), zero items are stored — zero cost.

Aggressiveness tiers (autoCapture.level):

| Level | What gets extracted | |-------|-------------------| | off | Nothing (disabled) | | facts | Only concrete facts and explicit decisions | | standard | Facts, decisions, preferences, and todos (default) | | full | Everything above plus a context summary embedding |

Flush triggers (whichever fires first):

| Trigger | Default | Config key | |---------|---------|------------| | User turn count | 10 turns | autoCapture.flushAfterTurns | | Idle timeout | 5 minutes | autoCapture.idleFlushMs | | Session end / shutdown | Automatic | — |

Killswitch: Add "conversations" to search.excludeBuckets to exclude auto-captured items from default brain_search results without deleting the data. You can still search them explicitly with brain_search(bucket: "conversations").

Example: A 10-turn debugging exchange produces 1-2 high-signal items (the root cause, the fix applied) instead of 10 low-value raw message embeddings.

The inbox entry is cleaned up after successful routing.

---

Under the Hood

Classification Pipeline

When you drop a thought, Brain sends it to an LLM (Claude or Gemini) with a classification prompt. The LLM returns structured JSON with:

• bucket: which of the 8 buckets it belongs to
• confidence: how sure the model is (0.0 to 1.0)
• title: a short label (≤8 words)
• summary: 1-2 sentence distillation
• nextActions: concrete next steps
• entities: extracted people, dates, amounts, locations
• urgency: now, today, this-week, or someday
• followUpDate: specific date if applicable
• tags: up to 3 auto-generated tags
• detectedIntent: reminder, todo, purchase, call, booking, or none

Drops

A drop is the fundamental unit of Brain. One thought = one drop. You don't need to think about where it goes, what to tag it, or how to format it. Just drop it.

Each drop is: 1. Saved to the inbox with a timestamp and embedding vector 2. Classified by an LLM into a bucket with a confidence score 3. Routed to the right bucket (or flagged for review if confidence is low) 4. Checked for duplicates and auto-merged if a near-match exists 5. Scanned for time-sensitive actions (reminders, bookings, etc.)

Photos are also supported. If you drop a photo and Tesseract is installed, Brain runs OCR to extract the text before classifying. If Tesseract isn't installed, the photo drop still works but skips text extraction silently, classifying based on any accompanying text instead.

Architecture

User Input
    │
    ▼
┌──────────┐    ┌────────────┐    ┌──────────┐    ┌──────────────┐
│ Tag Parse │───▶│  Embedder  │───▶│  Inbox   │───▶│  Classifier  │
│ [ToDo]…  │    │ (Gemini /  │    │ (LanceDB)│    │ (Claude /    │
└──────────┘    │  OpenAI)   │    └──────────┘    │  Gemini)     │
                └────────────┘                     └──────┬───────┘
                                                          │
                                                          ▼
                                                   ┌──────────────┐
                                                   │   Router     │
                                                   │ (Confidence  │
                                                   │  + Dedup)    │
                                                   └──────┬───────┘
                                                          │
                                          ┌───────────────┼───────────────┐
                                          ▼               ▼               ▼
                                    ┌──────────┐   ┌────────────┐  ┌─────────────┐
                                    │  Bucket  │   │   Needs    │  │ Auto-Merge  │
                                    │ (1 of 8) │   │   Review   │  │ (existing)  │
                                    └────┬─────┘   └────────────┘  └─────────────┘
                                         │
                                         ▼
                                  ┌──────────────┐
                                  │Action Router │
                                  │ (Reminders,  │
                                  │  ToDos, etc) │
                                  └──────────────┘

Key Technologies

• LanceDB: Embedded vector database for storage + semantic search (12 tables total)
• Embeddings: OpenAI text-embedding-3-small (1536 dim) or Gemini gemini-embedding-001 (3072 dim)
• Classification: LLM-powered via OpenClaw gateway (Claude) or direct Gemini API
• Reminders: Persistent cron jobs via openclaw cron with Telegram inline button delivery
• Audit Trail: Every action logged (capture, classify, route, merge, fix)

---

Reference

Agent Tools

| Tool | Parameters | Description | |------|-----------|-------------| | brain_drop | text (required), source, mediaPath | Capture a thought into the Brain | | brain_search | query (required), bucket, limit | Search by semantic similarity across all buckets or a specific one | | brain_stats | (none) | Show record counts and health for all buckets | | brain_audit | inputId, limit | View audit trail for a specific item or recent actions | | brain_digest | type (required): morning, midday, afternoon, night, weekly | Generate a digest (respects DND) | | brain_dnd | action (required): status, on, off | Control Do Not Disturb mode | | brain_fix | id (required), correction | Fix/move/trash/merge/inspect a Brain item |

brain_fix correction syntax:

| Syntax | What it does | |--------|-------------| | → people | Move item to the people bucket | | → trash | Delete item permanently | | action: "call them back" | Add/update the next action | | merge abc123 | Merge another item into this one | | (omitted) | Show item details and available fixes |

CLI Commands

| Command | Description | |---------|-------------| | openclaw brain stats | Show bucket record counts | | openclaw brain list <bucket> | List records in a bucket (e.g., openclaw brain list people --limit 10) | | openclaw brain drop "<text>" | Drop a thought from the command line | | openclaw brain audit | Show recent audit trail entries (--id <inputId> to filter, --limit <n>) |

---

Troubleshooting

"brain: no config provided, plugin inactive"

You need to add an embedding.apiKey to your Brain config. This is the only required field.

Drops stay in inbox (never classified)

Make sure classifier.apiKey is set. Without it, classification is disabled and drops remain in the inbox as "pending."

Items going to needs_review too often

Lower the classifier.confidenceThreshold (e.g., from 0.80 to 0.70). The default is conservative; a 70% threshold lets more items auto-route.

Duplicate items not being merged

Deduplication uses a cosine similarity threshold of 0.92. If two items are similar but not very similar, they'll be stored separately. You can manually merge with brain_fix(id: "abc", correction: "merge def").

Reminders not firing

Check that: 1. The action router config has a valid gatewayToken (same as classifier.apiKey) 2. The cron jobs exist: openclaw cron list 3. The reminder script exists at /home/clawdbot/clawd/scripts/brain-reminder.mjs

OCR not working on photo drops

Tesseract is optional and Brain works without it, but photo drops won't extract text. To enable OCR:

# Debian/Ubuntu
sudo apt install tesseract-ocr

# macOS
brew install tesseract

If Tesseract is missing, photo drops silently skip OCR and classify based on any text you include with the photo.

How to see everything Brain knows

• brain_stats: record counts per bucket
• brain_search("*"): browse items
• openclaw brain list <bucket>: list all records in a bucket
• brain_audit: view the full audit trail

---

Roadmap

Brain is actively evolving. Here's what's coming:

v2.1: Quality of Life

• [ ] Classification feedback messages: See where Brain routed your drop ("📂 Routed 'Call dentist' → admin, 93% confidence")
• [ ] Configurable digest schedule: Enable/disable individual digests, change times and word limits via config
• [ ] Needs-review workflow: Dedicated tools to triage low-confidence items quickly
• [ ] Search result details: Richer results with summaries and next actions in one call
• [ ] Bulk operations: Trash or move multiple items at once
• [ ] Voice note transcription: Whisper integration for automatic voice-to-text capture

v2.5: Intelligence

• [x] Payment actions: DOGE-powered actions via the wallet plugin. Drops like "send 5 DOGE to Alice" trigger recipient resolution, approval UI with inline buttons, and wallet_send execution with race-guard protection.
• [ ] Custom buckets: Add, remove, split, and merge buckets to fit your workflow. Define custom fields, classification behavior, and routing rules per bucket. Your Brain, your structure.
• [ ] Drops as work orders: Use drops to trigger dev_task (code/doc changes) and research (web search + summary) actions, with tiered gating for safety
• [x] ~~Payment actions~~: (Shipped in v2.5 — see above)
• [ ] Hybrid search: Combine keyword + semantic search for better results
• [ ] Cross-bucket links: Automatically detect and surface relationships between items (e.g., a person linked to a project)
• [ ] Learning from corrections: When you move a misclassified item, Brain learns and improves over time
• [ ] Few-shot classification: Build a bank of gold-standard examples from your corrections for better accuracy
• [ ] Smart merge suggestions: Prompt before creating near-duplicates
• [ ] LLM-powered weekly review: Natural language insights instead of template summaries

v3.0: Autonomous Memory

• [ ] Knowledge graph visualization: Interactive web view of your entire Brain as a connected graph
• [ ] Context-aware proactive recall: Brain surfaces relevant items during conversations without being asked
• [ ] Calendar integration: Cross-reference today's events with Brain items in morning digests
• [ ] Spaced repetition: Important items resurface at increasing intervals so you don't forget
• [ ] Location-aware reminders: "Remind me to buy milk when I'm near the store"
• [ ] Email capture: Forward emails to Brain for automatic extraction and classification

See the full roadmap for detailed descriptions, complexity estimates, and priority matrix.

---

License

MIT — Built by Quackstro LLC

---

Support the Project

If you find Brain useful, tips are always appreciated:

DOGE Address: D6i8TeepmrGztENxdME84d2x5UVjLWncat

Every DOGE goes toward hosting, continued development, and keeping the lights on. 🧠🐕