Sign in

Continual learning for LLMs.

Phronais enables LLMs to learn from interactions, extract reusable patterns, and apply knowledge across sessions.

Get startedLearn more
terminal
claude.json
# Add memory to Claude
claude mcp add phronais


# Done! Try it:
claude "remember my name"

Getting Started

Phronais - Continual Learning for LLMs

Give your AI persistent memory and reusable skills across all your sessions.

Installation

Get set up in under 60 seconds with Claude, ChatGPT, or any MCP client.

Core Concepts

Learn how memories and skills work together to build intelligence.

Custom Tools

Create your own MCP tools - HTTP proxies, transforms, and code runners.

API Reference

Complete reference for all 60+ MCP tools available.

LLMs are stateless, but they don't have to feel that way. Phronais is an MCP server that enables AI assistants to remember your preferences, learn from interactions, and apply accumulated knowledge across sessions and devices. Works with Claude, ChatGPT, and any MCP-compatible client.

Quick Start

Claude Code CLI

Connect Phronais in one command:

claude mcp add phronais --transport http --scope user https://mcp.phronais.com/mcp

Click the authentication link when prompted, sign in with Google, and you're ready to go.

ChatGPT

Add Phronais as an MCP server with URL https://mcp.phronais.com/mcp and sign in with Google.

Your First Memory

Store observations about your work that your AI can recall later:

"remember that my API endpoints are in src/routes/ and use Hono framework"
"remember that I prefer async/await over .then() syntax"
"remember that database migrations use Drizzle ORM"

Build Skills from Experience

After storing memories, detect patterns and codify them as reusable skills:

analyze patterns in domain "cloudflare"
→ "Detected pattern: You always check CORS headers first when debugging API calls"

generate skill from pattern
→ Creates "API Debugging Checklist" starting with CORS verification

approve skill
→ Now active and suggested automatically when debugging

User Approval Required

Skills always start as drafts and require your explicit approval before activation. This prevents your AI from auto-learning unwanted behaviors.

Key Features

60+ MCP Tools

Phronais provides a comprehensive toolkit organized into categories:

CategoryToolsWhat They Do
Memory (4)remember, recall, forget, correctStore and retrieve observations
Skills (14)create, update, approve, deprecate, rollback, searchManage reusable knowledge patterns
Learning (5)analyze_patterns, generate_skill, suggest_skillAI-assisted pattern detection
Dynamic Tools (7)create, update, delete, list, get, testBuild custom MCP tools
Code Tools (4)create_code_tool, approve, reject, list_pendingRun JavaScript in isolated containers
Agent Coordination (13)heartbeat, create_task, claim_task, send_messageMulti-agent collaboration

Create Custom Tools

Turn frequent API calls into reusable MCP tools:

// HTTP Proxy - call any API
create_dynamic_tool({
  name: "github_create_issue",
  tool_type: "http_proxy",
  http_config: {
    method: "POST",
    url_template: "https://api.github.com/repos/{{owner}}/{{repo}}/issues",
    connection_id: "my-github-connection"
  }
})

Code Tools with Sandbox Isolation

Run custom JavaScript in isolated Cloudflare Sandbox containers:

create_code_tool({
  name: "calculate_stats",
  runtime: "javascript",
  code: `
    export default {
      async fetch(request) {
        const { numbers } = await request.json();
        const sum = numbers.reduce((a, b) => a + b, 0);
        return Response.json({
          sum,
          average: sum / numbers.length
        });
      }
    }
  `
})

Cross-Device and Cross-Client Sync

Sign in with the same Google account on any device or client - all your memories and skills sync automatically.

Claude Desktop:  Store 50 memories about your codebase
ChatGPT:         All 50 memories immediately available
Claude Code:     Same data, same knowledge

How It Works

1. Store Memories

During interactions, store observations about what worked:

remember that CORS issue was fixed by adding Access-Control-Expose-Headers
→ Stored with context: domain=cloudflare, outcome=success

2. Detect Patterns

Analyze memories to find repeated approaches:

analyze patterns in domain "cloudflare" from last 30 days
→ Found: 3 memories about CORS debugging with 0.85 confidence

3. Generate Skills

Convert patterns into reusable skills:

generate skill from memories about CORS fixes
→ Creates draft: "Cloudflare Workers CORS Configuration"

4. Approve & Apply

Review and approve skills before activation:

list_skills status "draft"
→ Shows pending skills awaiting approval

approve skill "cloudflare-workers-cors"
→ Skill now active, suggested automatically in relevant contexts

Getting Help

Next
Installation