Extension Catalog

Extensions for the-brain — swap any layer, harvest from any AI tool, train however you want. Mix and match.

The extensions below were the project's research starter pack. This package list is preserved for reference; community extension development is no longer active.

🔍 Browse by Layer

Filter packages by the cognitive layer they operate on.

• ⚡ Instant
• ⚖️ Selection
• 🌌 Deep
• 📥 Harvesters
• 🔧 Infrastructure

⚡ Instant Layer

Real-time context injection — corrections, preferences, and patterns detected immediately.

@the-brain-dev/plugin-graph-memory

• Version: v0.1.0
• Type: Memory
• Status: Starter pack
• Description: Creates fast relational nodes of recent corrections. Language-agnostic structural heuristics with weight boost on match and weight decay for stale nodes.

⚖️ Selection Layer

The Gatekeeper — filters noise from signal. Only surprising, high-value memories are promoted to deep storage.

@the-brain-dev/plugin-spm-curator

• Version: v0.1.0
• Type: Memory
• Status: Starter pack
• Description: Surprise-Gated Prediction Error filter. Composite surprise score using scalar, embedding, and novelty metrics. Configurable threshold (default: 0.30).

🌌 Deep Layer

Permanent consolidation — overnight LoRA training, identity anchoring, and knowledge wiki generation.

@the-brain-dev/trainer-local-mlx

• Version: v0.1.0
• Type: Trainer
• Status: Starter pack (macOS only)
• Description: Runs overnight at 2 AM. Trains LoRA adapters on consolidated DEEP memories. Fully local — data never leaves your machine. Requires Apple Silicon + uv.

@the-brain-dev/plugin-identity-anchor

• Version: v0.1.0
• Type: Memory
• Status: Starter pack
• Description: Maintains a stable self-vector across LoRA retrains. Detects drift in coding style and preferences.

@the-brain-dev/plugin-auto-wiki

• Version: v0.1.0
• Type: Output
• Status: Starter pack
• Description: Weekly static wiki output from consolidated knowledge. Interlinked markdown pages with concept nodes and entity pages.

📥 Harvesters

Harvesters — collect interactions from any AI tool and push them into the pipeline.

@the-brain-dev/plugin-harvester-cursor

• Version: v0.1.0
• Type: Harvester
• Status: Starter pack
• Description: Reads Cursor's local SQLite workspace database. Auto-detects project context via workDir matching.

@the-brain-dev/plugin-harvester-claude

• Version: v0.1.0
• Type: Harvester
• Status: Starter pack
• Description: Reads Claude Code's local JSONL conversation logs. Tracks offset/checkpoint to avoid re-processing.

@the-brain-dev/plugin-harvester-gemini

• Version: v0.1.0
• Type: Harvester
• Status: Starter pack
• Description: Polls Gemini CLI's local JSON logs for new interactions.

@the-brain-dev/plugin-harvester-hermes

• Version: v0.1.0
• Type: Harvester
• Status: Starter pack
• Description: Polls Hermes Agent's SQLite state.db for user↔assistant chat interactions.

@the-brain-dev/plugin-harvester-lm-eval

• Version: v0.1.0
• Type: Harvester
• Status: Starter pack
• Description: Harvests lm-evaluation-harness benchmark results. Computes per-model per-benchmark fingerprints with anomaly detection (>2σ). Enables the-brain as cognitive layer for meta-harness systems (AHE, Meta-Harness).

🔧 Infrastructure

Core framework, integrations, and storage backends.

@the-brain-dev/core

• Version: v0.1.0
• Type: Core
• Status: Starter pack
• Description: Empty data bus — everything else is a plugin via definePlugin(). Includes hook system, BrainDB (Drizzle ORM + sqlite), LayerRouter, ProjectManager, ContentCleaner, AuthDB (multi-user auth), and PermissionResolver.

@the-brain-dev/mcp-server

• Version: v0.1.0
• Type: Integration
• Status: Starter pack
• Description: MCP (Model Context Protocol) server — exposes the-brain as a memory backend for Claude Desktop, Cursor, Zed. 26 tools + 11 resources + SSE subscriptions + meta-harness integration tools.

@the-brain-dev/storage-libsql

• Version: v0.1.0
• Type: Storage
• Status: Starter pack (optional)
• Description: Alternative to SQLite. Enables shared database across multiple servers for team deployments via Turso.

Architecture Overview

📥 Harvesters  →  ⚡ Instant Layer  →  ⚖️ Selection Layer  →  🌌 Deep Layer  →  🔌 Clients

cursor          graph-memory         spm-curator             trainer-local-mlx     MCP Server
claude                                (surprise filter)      identity-anchor       Claude Desktop
gemini                                                       auto-wiki             Cursor, Zed
hermes                                                       lm-eval harvester     AHE, Meta-Harness
lm-eval                                                      (benchmark data)      (via MCP tools)

Building Your Own Plugin

See Writing Plugins for the full plugin authoring guide.

Extension Commands

Extensions register custom commands accessible via the-brain ext <cmd>:

# List all registered extension commands
the-brain ext --list

⚠️ Limitation: brain.registerCommand() registers commands in the daemon process memory. The CLI process (the-brain ext) has its own copy and can't see daemon-registered commands. For extensions that need CLI access, use a standalone runner script (see Writing Plugins for the pattern).