RecallX

Your AI's Memory, Across Every Coding Session.

Universal local-first persistent memory for AI coding agents. Remember project context, technical decisions, debugging solutions, and developer preferences across sessions.

The Problem

AI coding agents lose context between sessions. Every new conversation starts from scratch. You repeatedly explain your architecture, preferences, decisions, and previous work. Context is lost, productivity suffers.

Related MCP server: memento

The Solution

RecallX provides a persistent, searchable, structured memory layer that any AI coding agent can access through the Model Context Protocol (MCP). Your agent remembers what matters — automatically.

Key Features

• Local-first — All data stays on your machine. No cloud, no telemetry, no external APIs.

• Agent-independent — Works with Claude Code, Cursor, Gemini CLI, Codex, OpenCode, and any MCP-compatible agent.

• Project-aware — Memories are isolated by project. No cross-contamination.

• Intelligent retrieval — Full-text search with relevance ranking. Optional semantic search with local embeddings.

• Structured memories — Project context, technical decisions, bug solutions, developer preferences, session summaries.

• Privacy-first — Automatic secret detection and redaction. No hidden network requests.

• CLI + MCP — Full-featured command line interface and Model Context Protocol server.

Quick Start

Installation

pip install recallx

Or with pipx for isolated installation:

pipx install recallx

Initialize

recallx init

Save a Memory

recallx remember "Using PostgreSQL with SQLAlchemy for the API backend" -p my-api -t decision --tags database,architecture -i 0.8

Search Memories

recallx search "database architecture" -p my-api

Check Status

recallx status

MCP Integration

RecallX integrates with AI coding agents through the Model Context Protocol.

Claude Code

Add to your Claude Code MCP settings:

{
  "mcpServers": {
    "recallx": {
      "command": "recallx",
      "args": ["mcp", "serve"]
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "recallx": {
      "command": "recallx",
      "args": ["mcp", "serve"]
    }
  }
}

Gemini CLI / Antigravity

Add to .gemini/settings.json:

{
  "mcpServers": {
    "recallx": {
      "command": "recallx",
      "args": ["mcp", "serve"]
    }
  }
}

Run recallx integrate <agent> for specific setup instructions.

CLI Reference

MCP Tools

When connected via MCP, RecallX exposes these tools to your AI agent:

Agent Skills

RecallX includes reusable agent skills for structured memory operations:

• recallx-remember — Guidelines for identifying and saving durable information

• recallx-recall — Guidelines for retrieving relevant context before starting work

• recallx-project-context — Guidelines for maintaining project context

Skills are located in src/recallx/skills/ and can be copied to your agent's skills directory.

Memory Types

Storage and Privacy

• All data stored locally in SQLite

• Default location:

  • Windows: %LOCALAPPDATA%\RecallX\

  • Linux: ~/.local/share/recallx/

  • macOS: ~/Library/Application Support/RecallX/

• No telemetry, analytics, or external data transmission

• Automatic detection and redaction of secrets (API keys, tokens, passwords)

• Full export/import for data portability

Optional Semantic Search

For enhanced retrieval, install with semantic search support:

pip install recallx[semantic]

This downloads a local embedding model (~90MB). No external API required. RecallX works perfectly without it — FTS5 full-text search handles most use cases.

Architecture

┌─────────────────────────────────────────────────┐
│                  AI Coding Agent                │
│         (Claude, Cursor, Gemini, etc.)          │
└──────────────────────┬──────────────────────────┘
                       │ MCP (stdio)
┌──────────────────────▼──────────────────────────┐
│                  RecallX MCP Server             │
│              Tools · Resources · Prompts        │
├─────────────────────────────────────────────────┤
│                  Memory Engine                  │
│     Extraction · Retrieval · Ranking · Dedup    │
├─────────────────────────────────────────────────┤
│                  Search Layer                   │
│         FTS5 · Embeddings · Hybrid Search       │
├─────────────────────────────────────────────────┤
│                  Storage Layer                  │
│           SQLite · Repositories · Migrations    │
└─────────────────────────────────────────────────┘

Limitations

• Semantic search requires additional disk space for the embedding model

• Memory extraction uses rule-based patterns (no LLM-powered extraction in v1)

• MCP servers cannot automatically access agent conversation histories

• Performance depends on database size (optimized for up to ~100K memories)

Roadmap

• Web-based memory browser (optional)

• LLM-powered memory extraction

• Memory graph visualization

• Team/shared memory (opt-in)

• Plugin system for custom memory types

• Automatic session capture hooks

Contributing

See CONTRIBUTING.md for development setup and guidelines.

Security

See SECURITY.md for security policies and reporting.

License

MIT License. See LICENSE for details.

RecallX — Built with engineering discipline, privacy, and reliability as the highest priorities.