MUR - Continuous learning for AI assistants

MUR is a local-first platform for long-lived AI agent fleets, evolving YAML-based memory, workflows, permissions, and knowledge synchronization across supported AI tools.

Visit Website
MUR - Continuous learning for AI assistants

Introduction

Overview

MUR is a local-first platform for running specialized AI agents as long-lived processes. A small Rust runtime supervises agents with their own model bindings, prompts, MCP servers, skills, schedules, and least-privilege permissions.

Knowledge learned during work is captured, scored, stored as plain YAML, and made available in later sessions. This gives multiple AI assistants a shared memory while keeping the data on the user's computer.

Key Features

  • Maintains a skill lifecycle from Draft through Emerging, Stable, and Canonical, with decay for knowledge that is no longer useful.
  • Synchronizes one knowledge base with more than 16 AI tools, including Claude Code, Gemini CLI, Cursor, Copilot CLI, Codex, Aider, Windsurf, and OpenCode.
  • Combines vector similarity, BM25 keyword matching, and multi-factor scoring for hybrid semantic search.
  • Runs agent fleets and workflows with deadlines, cost limits, and a kill switch.
  • Keeps skills in human-readable YAML suitable for version control.
  • Supports durable monitoring of CI runs, fleet runs, and subprocesses across daemon restarts.

Agent Workflows

MUR can run a main coding agent and a reviewing agent in repeated rounds until the reviewer approves, blocks, or reaches a configured limit. Review findings receive stable IDs, disputed findings can wait for a user decision, and interrupted sessions can resume after the last completed round while preserving the channel as an audit record.

Deep research distributes a question among multiple workers under a defined budget and saves a report. Background model routing considers whether a candidate model supports the task's requirements, while user-selected models remain unchanged unless smart routing is enabled.

Control and Privacy

Actions that reach outside MUR pass through a risk gate and can pause for approval without timing out into automatic execution. Approvals are tied to the content of an action, explicit rejection overrides standing policy, and automated approvals are recorded.

MUR stores data locally with zero telemetry; cloud synchronization is optional. It can be installed on macOS, Linux, and Windows, and a new agent can default to a local Ollama model without requiring an API key.