claudecode-rlm

Unlimited context for Claude Code - Knowledge graph-based context storage with automatic memory.

Author: Michael Thornton ([email protected]) Repository: https://github.com/tekcin/claudecode-rlm

Features

• TRUE Unlimited Context - Proxy server auto-stores ALL conversations, auto-injects relevant context
• MCP Server - Manual memory tools for Claude Code
• Knowledge Graph - Hierarchical storage (Document → Section → Chunk → Entity)
• 74x Faster Reads - LRU cache with inverted index
• Entity Extraction - Automatic extraction of code elements, files, concepts
• Recency Weighting - Fresh context ranked higher

Installation

npm install -g claudecode-rlm

🚀 Proxy Mode (Recommended) - TRUE Automatic Memory

The proxy server intercepts ALL API calls for complete automatic context:

1. Start the proxy

claudecode-rlm-proxy

2. Configure Claude Code

export ANTHROPIC_BASE_URL=http://localhost:3456

Or add to your shell profile (~/.bashrc, ~/.zshrc):

export ANTHROPIC_BASE_URL=http://localhost:3456

What happens automatically:

• ✅ Every user message → Stored in knowledge graph
• ✅ Every assistant response → Stored in knowledge graph
• ✅ Before each request → Relevant past context injected
• ✅ Streaming supported → Works with Claude Code's streaming

Proxy Environment Variables

| Variable | Description | Default | |----------|-------------|---------| | CLAUDECODE_RLM_PROXY_PORT | Proxy port | 3456 | | CLAUDECODE_RLM_WORKDIR | Storage directory | Current directory | | CLAUDECODE_RLM_SESSION | Session ID | auto | | CLAUDECODE_RLM_MAX_INJECT | Max tokens to inject | 2000 | | ANTHROPIC_REAL_URL | Real Anthropic API URL | https://api.anthropic.com |

MCP Mode - Manual Memory Tools

For manual control via MCP tools:

Add to Claude Code config (~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "claudecode-rlm": {
      "command": "claudecode-rlm"
    }
  }
}

Available Tools

| Tool | Description | |------|-------------| | memory_auto | Store AND retrieve context in one call (recommended) | | memory_store | Store content in knowledge graph | | memory_search | Search with recency weighting | | graph_query | Query graph (search/entity/expand/path) | | list_entities | List tracked entities | | graph_stats | Get storage statistics |

Architecture

Knowledge Graph Structure

Document (conversation turn)
├── Section (topic/header group)
│   ├── Chunk (~300 chars, searchable)
│   │   ├── Entity (function name)
│   │   ├── Entity (file path)
│   │   └── Entity (concept)
│   └── Chunk
│       └── [FOLLOWS] → next chunk
└── Section

Storage Location

.claude/claudecode-rlm/graph/{sessionID}/
├── nodes/*.json    - Content nodes
├── edges/*.json    - Relationships
└── indexes/*.json  - Search indexes

Performance

| Operation | Basic | Optimized | Speedup | |-----------|-------|-----------|---------| | Node reads | 21.3µs | 0.3µs | 74x | | Get by type | 1.1ms | 21µs | 54x | | Bulk insert | 6.07s | 1.21s | 5x |

Library Usage

import { GraphStorage, GraphIngester, EnhancedSearch } from 'claudecode-rlm'

GraphStorage.init('/path/to/project')

// Store
GraphIngester.ingestContextBlock({
  id: 'block-1',
  sessionID: 'my-session',
  content: 'Your content...',
  summary: '',
  tokens: 1000,
  createdAt: Date.now()
})

// Search
const results = EnhancedSearch.search('my-session', 'query')

Development

npm install
npm run build
npm start           # MCP server
npm run proxy       # Proxy server

License

MIT

Author

Michael Thornton

• Email: [email protected]
• GitHub: @tekcin

Links

• npm package
• GitHub
• Issues