lodis

Lodis

Universal, portable memory layer for AI agents.

Lodis gives your AI tools a shared memory — searchable, correctable, and under your control. Install once, connect to Claude Code, Cursor, Windsurf, or any MCP-compatible client.

Quick Start

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

{
  "mcpServers": {
    "lodis": {
      "command": "npx",
      "args": ["-y", "@sunriselabs/lodis"]
    }
  }
}

That's it. Your AI now has persistent memory.

Migrating from engrams? This project was published as engrams on npm prior to v0.6.0. The package is now @sunriselabs/lodis — same code, same data directory (~/.lodis/), same MCP tools. To migrate, swap engrams for @sunriselabs/lodis in your MCP config ("args": ["-y", "@sunriselabs/lodis"]) and reinstall. The old engrams package on npm is frozen at v0.5.1 and will not receive further updates.

Claude Code Plugin

Prefer a one-step install that brings the MCP server and the memory skills together? Lodis ships as a Claude Code plugin:

/plugin marketplace add Sunrise-Labs-Dot-AI/lodis
/plugin install lodis@lodis-official

Installing the plugin registers the lodis MCP server (no manual config edit) and adds four skills that teach the agent to use memory well:

Skill	What it does
/lodis:memory-retrieval	Drive memory_context at session start with adaptive budgets, then rate the retrieval
/lodis:memory-capture	Write discipline — durable fact vs. progress snippet, dedup resolution, entity typing, connections
/lodis:onboarding	First-run setup: configure the agent to prefer Lodis, scan tools, interview, seed memories
/lodis:session-wrap	End-of-session distillation into durable memory before context is lost

The plugin manifest lives in .claude-plugin/; the skills are in skills/.

Using Codex, Cursor, Cline, or another MCP client?

The plugin is Claude-Code-only, but the MCP server and all 40 tools work everywhere. To give those agents the same memory know-how the skills provide, connect the server (e.g. Codex: codex mcp add lodis -- npx -y @sunriselabs/lodis) and paste the Lodis memory policy into your client's instruction file (AGENTS.md, .cursor/rules, Cline custom instructions, etc.).

Getting Started

After installing, tell your AI assistant:

"Help me set up Lodis"

The assistant will call memory_onboard and:

1. Scan your connected tools (calendar, email, GitHub) to extract people, projects, and context
2. Interview you with targeted questions based on what it found
3. Seed 30-50 memories with entity types and connections

Review your memories at localhost:3838. Confirm what's right, correct what's wrong.

Importing Existing Memories

If you have memories in other tools, your AI can import them:

• Claude Code auto-memory: "Import my Claude memories into Lodis"
• ChatGPT memory export: "Import this ChatGPT memory export into Lodis"
• Cursor rules: "Import my .cursorrules as Lodis preferences"

The memory_import tool handles parsing and deduplication. Where semantic judgment is needed, the calling agent supplies entity fields or follows up with memory_classify / memory_update.

What It Does

• Remembers across tools. Teach Claude something, Cursor knows it too.
• Searches semantically. Hybrid search (full-text + vector embeddings) finds relevant memories even with different wording.
• Knows what it knows. Confidence scoring, source attribution, and entity classification on every memory.
• Lets you correct it. Confirm, correct, split, or remove memories through the dashboard or MCP tools.
• Deduplicates on write. Similar memories are detected and surfaced to the agent for resolution.
• Builds a knowledge graph. Memories connect to each other with typed relationships. 14 entity types (people, organizations, projects, preferences, snippets, and more) can be supplied by agents and linked by deterministic helpers or agent-reviewed connection proposals.
• Manages memory permanence. Four tiers — canonical, active, ephemeral (TTL), and archived — control confidence decay and search ranking.
• Packs context efficiently. Token-budget-aware retrieval via memory_context delivers the right amount of context for any LLM window.
• Generates entity profiles. On-demand summaries of known people, projects, and organizations via memory_briefing.
• Indexes external documents. Pull in context from Google Drive, Notion, or local files for unified search.