mathomhaus/guild

What it solves

Guild addresses the problem of "amnesiac" AI coding agents that lose their context, memory, and progress when a session ends or the context window is exhausted. It provides a persistent, shared state that allows multiple agents—even those using different MCP clients like Claude Code, Cursor, or Codex—to coordinate tasks and share knowledge without duplicating effort or colliding on the same work.

How it works

Guild operates as a single Go binary and an MCP server using an embedded SQLite database to store state locally. It uses four primary primitives to manage the agent lifecycle:

• Quests: Task tracking with priorities, dependencies, and atomic locks to prevent multiple agents from claiming the same task.
• Lore: A knowledge archive for observations, decisions, and research, searchable via a hybrid approach combining keyword (BM25) and vector similarity (reciprocal-rank fusion).
• Oaths: A set of core principles (lore with kind=principle) that are automatically loaded at the start of every session.
• Briefs: Handoff notes written by one agent to inform the next session's agent.

Who it’s for

Developers using MCP-enabled AI coding agents (such as Claude Code, Cursor, or Codex) who want to maintain long-term project memory and coordinate complex tasks across different sessions or parallel agent instances.

Highlights

• Local-first state: All data stays on the local host in SQLite; no external API keys or accounts are required.
• Hybrid Search: Combines lexical BM25 and vector similarity for semantic retrieval of project knowledge.
• Atomic Task Coordination: Prevents collisions through atomic quest claims and manages task dependencies via cascading unblocks.
• Client Agnostic: Works across any MCP-compatible editor or agent harness.