🚀 v0.0.20 🚀

Fiber-based concurrency is now the default.
Parallel embedding, tagging, and proposition extraction via simple_flow.

HTM

Give your AI tools a shared, persistent memory.

Your AI assistants are brilliant but forgetful. Claude Desktop doesn't remember what Claude Code learned. Your custom chatbot can't recall what happened yesterday. Every tool starts fresh, every session.

HTM fixes this.

Connect all your MCP-enabled tools to the same long-term memory. What one robot learns, all robots remember. Context persists across sessions, tools, and time.

• Hierarchical: Organizes information from simple concepts to detailed relationships
• Temporal: Retrieves memories across time—from moments ago to months past
• Memory: Encodes, stores, and recalls information with semantic understanding

HTM is a Ruby gem that provides durable, shared memory for LLM-powered applications.

HTM includes a Model Context Protocol server with 23 tools for memory management. Connect Claude Desktop, Claude Code, AIA, or any MCP-compatible client:

{
  "mcpServers": {
    "htm": {
      "command": "htm_mcp",
      "env": { "HTM_DATABASE__URL": "postgresql://localhost/htm" }
    }
  }
}

The htm_mcp executable includes CLI commands for database management:

htm_mcp setup    # Initialize database schema
htm_mcp verify   # Check connection and migrations
htm_mcp stats    # Show memory statistics
htm_mcp help     # Full help with environment variables
htm_mcp          # Start MCP server (default)

Now your AI assistant can remember and recall across sessions:

• Store decisions, preferences, and context
• Search memories with natural language
• Build knowledge that compounds over time

All robots share the same global memory. What one learns, all can access:

# Claude Desktop remembers a user preference
claude_desktop.remember("User prefers dark mode and concise responses")

# Later, Claude Code can recall it
claude_code.recall("user preferences")
# => "User prefers dark mode and concise responses"

Every tool builds on what came before. No more repeating yourself.

Two-tier architecture ensures nothing is lost:

• Working Memory: Token-limited context for immediate use
• Long-Term Memory: Durable PostgreSQL storage with vector search

Memories persist forever unless explicitly deleted. Working memory evicts to long-term storage—it never deletes.

htm.remember("PostgreSQL with pgvector handles our vector search")

# Months later...
htm.recall("database decisions", timeframe: "last year")
# => Still there

Coordinate multiple robots with shared working memory and instant failover:

group = HTM::RobotGroup.new(
  name: 'support-ha',
  active: ['primary'],
  passive: ['standby']
)

# Primary fails? Standby takes over instantly
group.failover!

Real-time synchronization via PostgreSQL LISTEN/NOTIFY. Add or remove robots dynamically.

gem install htm

require 'htm'

# Initialize
htm = HTM.new(robot_name: "MyAssistant")

# Remember
htm.remember("User's project uses Rails 7 with PostgreSQL")

# Recall
memories = htm.recall("tech stack", timeframe: "last month")

For MCP integration, database setup, and configuration options, see the full documentation.

• Ruby 3.0+
• PostgreSQL with pgvector and pg_trgm extensions
• Ollama (default) or any local/private provider for embeddings and classifications

https://madbomber.github.io/htm

• Installation & Setup
• MCP Server Guide
• Rails Integration
• API Reference

"What's in a name? That which we call a rose By any other name would smell as sweet." — Shakespeare, Romeo and Juliet

Shakespeare argues names are arbitrary. In software, we respectfully disagree—names shape expectations and understanding.

HTM uses "robots" rather than the fashionable "agents" for several reasons:

• Semantic clarity: "Agent" is overloaded—user agents, software agents, real estate agents, secret agents. "Robot" is specific to automated workers.

• Honest about capabilities: "Agent" implies autonomy and genuine decision-making. These systems follow instructions—they don't have agency. "Robot" acknowledges what they actually are: tools.

• Avoiding the hype cycle: "AI Agent" and "Agentic AI" became buzzwords in 2023-2024, often meaning nothing more than "LLM with a prompt." We prefer terminology that will age well.

• Many "agent" frameworks are prompt wrappers: Look under the hood of popular agent libraries and you'll often find a single prompt in a for-loop. Calling that an "agent" sets false expectations.

• Rich heritage: "Robot" comes from Karel Čapek's 1920 play R.U.R. (from Czech robota, meaning labor). Isaac Asimov gave us the Three Laws. There's decades of thoughtful writing about robot ethics and behavior. "Agent" has no comparable tradition.

• Personality: Robots are endearing—R2-D2, Wall-E, Bender. They have cultural weight that "agent" lacks.

Honest terminology leads to clearer thinking. These are robots: tireless workers with perfect memory, executing instructions on our behalf. That's exactly what HTM helps them do better.

Bug reports and pull requests are welcome on GitHub at https://github.com/madbomber/htm.

The gem is available as open source under the terms of the MIT License.