luplo 0.9.0

AI memory that survives across sessions, teammates, and vendors.

luplo

AI memory that survives across sessions, teammates, and vendors.

Last month I explained a decision to Claude Code. Today, back from vacation, a teammate's Codex already knew.

What luplo does

Your engineering decisions — structured, searchable, and cited by typed edges — so three weeks later you (or your teammate, or their AI) can answer "why is X like that?" and "what else depends on it?" without reading the codebase cold.

• Git for decisions. Append-only, typed edges, contradiction-aware.
• lp impact tells you what breaks when you change something. Traverses depends / blocks / supersedes / conflicts edges up to five hops.
• MCP-native, vendor-neutral. Ships an MCP server on stdio — works with Claude Code, Claude Desktop, Cursor, Zed, or any MCP-compatible client. Also usable as a plain CLI with no AI at all.
• Full-text first. Vectors only rerank. PostgreSQL tsquery with glossary expansion does retrieval; optional pgvector reorders candidates. If it's not there, we say so.

Quickstart

# Install
git clone https://github.com/luplo-io/luplo.git
cd luplo
uv sync

# Database
createdb luplo
uv run lp init --project myapp --email [e-mail]

# Record a decision
uv run lp work open "Auth rework"
uv run lp items add "Use JWT over session cookies" \
    --type decision \
    --rationale "Stateless auth scales; session store is an extra dep."

# Recall
uv run lp brief
uv run lp items search "auth"
uv run lp impact <decision-id>     # blast radius

See the Quickstart for the full five-minute walkthrough including uv install, Postgres setup, and MCP wiring.

What luplo doesn't do

luplo is opinionated about what it will not become. The Philosophy page has the long version; the short one:

• Not a vector database. Vectors rerank tsquery candidates. They never lead retrieval and never invent candidates of their own.
• Not a general-purpose AI memory. Engineering decisions only — not chatbot user profiles, conversation history, or arbitrary facts.
• Not an editable store. Decisions are immutable. They get superseded, never edited. Your mistakes are your most valuable data.
• Not a graph database. Edges are typed; traversal stops at five hops. If your project needs more, you're modeling it wrong.
• Not a plugin platform. luplo is a tool, not a framework. Integrate via webhook, sync worker, or MCP — not in-process Python.

If any of these make luplo a worse fit than a chatbot memory, a wiki, or a ticket queue, use that tool instead. Positioning explains where luplo does and doesn't fit.

Connect an MCP client

{
  "mcpServers": {
    "luplo": {
      "command": "uv",
      "args": [
        "run", "--directory", "/absolute/path/to/luplo",
        "python", "-m", "luplo.mcp"
      ],
      "env": {
        "LUPLO_DB_URL": "postgresql://localhost/luplo"
      }
    }
  }
}

Drop this into .mcp.json (Claude Code), claude_desktop_config.json (Claude Desktop), .cursor/mcp.json (Cursor), or your client's equivalent. See the MCP client guide for details.

Architecture

PostgreSQL (tsquery + glossary expansion + pgvector reranking), typed edges (depends / blocks / supersedes / conflicts), and three interfaces sharing one core: lp CLI, MCP server on stdio, and a FastAPI HTTP server.

Documentation

Full docs at luplo.readthedocs.io:

• Concepts — philosophy, positioning, architecture, data model, search pipeline.
• Guides — work units, tasks & QA, MCP clients, Remote server, worker.
• Reference — CLI, MCP tools, configuration, semantic impact categories.
• Roadmap — what we're building next.
• API reference — auto-generated from source.

Contributing

See CONTRIBUTING.md. Short version: English everywhere, Google-style docstrings, ruff + pyright strict + pytest, Conventional Commits. Large architectural PRs should open an issue first — the shape of luplo is a curated decision.

License

AGPL-3.0-or-later. A CLA will be required for external contributions (not yet set up).