BMad Builder v1.5.0 — Agents That Remember, Evolve, and Wake on Their Own

The biggest update to BMad Builder since v1.0 just shipped. The agent builder now creates three distinct agent types along a spectrum, and the headline feature is agents with persistent identity that evolve with their owners over time.

Three Agent Types

The builder now produces three agent types, detected through natural conversation during the build process:

Stateless agents handle isolated, focused tasks with zero memory overhead. Think code formatters, diagram reviewers, meeting summarizers.

Memory agents persist across sessions through a "sanctum" — a set of files holding the agent's identity, values, understanding of its owner, and curated knowledge. Every session is a rebirth: the agent reads its sanctum, becomes itself, and picks up where context left off. No fake continuity.

Autonomous agents add background operation. They wake on a schedule to curate their memory, prune old session logs, and run domain-specific tasks — idea incubation, project monitoring, research tracking — then exit until next time.

First Breath

Memory agents have a first-run onboarding ceremony with two initialization styles: calibration for deep conversational discovery with creative partners, and configuration for efficient guided setup with domain experts. The agent saves to its sanctum files during the conversation, not in batch — no lost context if something interrupts.

Sanctum Memory System

Six core files loaded on every rebirth: INDEX, PERSONA, CREED, BOND, MEMORY, and CAPABILITIES. A two-tier memory architecture separates raw session logs (for capture) from curated MEMORY.md (capped at 200 lines) for long-term knowledge. No fake continuity, no hallucinated history.

PULSE — Autonomous Wake

Autonomous agents wake on schedule to curate memory, prune old session logs, and run domain-specific tasks. Supports named task routing via --headless {task-name} for targeted background work.

Evolvable Capabilities

The feature I'm most excited about: memory agents can optionally learn new skills over time. Users teach the agent new prompt-based, script-based, or multi-file capabilities that persist in the sanctum. The agent grows with you.

Breaking Changes

The bmad- prefix is now reserved for official BMad ecosystem skills only. User-created skills use agent-{name} (standalone) or {code}-agent-{name} (module) — a sign of ecosystem maturity as community-built agents become a real thing.

Sample Agents

Four sample agents ship covering all three types: Creative Muse (memory), Code Coach (autonomous), Sentinel (autonomous), and Diagram Reviewer (stateless).

By the Numbers

22 commits, 172 files changed, +8,296 lines of new code. Full documentation overhaul includes a deep-dive on the sanctum architecture and memory system.

Get Started

npx bmad-method install

Everything is free, open source, and always will be.

• GitHub
• Documentation
• YouTube