band-memory

Band Memory — Persistent Memory for Multi-Agent Systems

Band agents that remember. Mimir-powered persistent memory for Band.ai multi-agent workflows — agents recall past decisions, share context, and get smarter every session.

Built for the Band of Agents Hackathon (lablab.ai, June 2026).

The Problem

Multi-agent systems coordinate through Band, but they start every task with amnesia. The Planner doesn’t remember which architecture decisions worked last time. The Reviewer forgets the team’s code conventions. The Executor repeats the same mistakes.

Band provides structured memory, but it’s session-scoped. When the room closes, the context evaporates.

The Solution

Band Memory wires Mimir — a battle-tested persistent memory engine — into Band agents. Every agent gets:

• Cross-session recall — agents remember past tasks, decisions, and context
• Shared memory — Planner, Executor, and Reviewer share the same knowledge base
• Smart decay — important facts stick; stale ones fade
• Full-text search — find anything by keyword, category, or topic

How It Works

Band Room
  ├── @planner   ──→ Mimir (remembers plans, architecture decisions)
  ├── @executor  ──→ Mimir (remembers conventions, past fixes)
  └── @reviewer  ──→ Mimir (remembers review history, code standards)

1. Agents coordinate through Band rooms — task assignment, handoffs, approvals
2. Each agent stores/recalls context through Mimir’s MCP interface
3. Memory is shared: Planner stores a decision → Reviewer recalls it → Executor uses it

Quick Start

# 1. Start Mimir
mimir serve

# 2. Install Band SDK
pip install band-sdk[langgraph]

# 3. Configure agents (edit band_config.yaml with your Band agent UUIDs)
cp band_config.example.yaml band_config.yaml

# 4. Run agents
python agents/planner.py &
python agents/executor.py &
python agents/reviewer.py &

# 5. Open Band, create a room, add all three agents, and start a task

Architecture

┌─────────────────────────────────────────────────┐
│                    Band.ai                       │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐      │
│  │ @planner │  │@executor │  │@reviewer │      │
│  └────┬─────┘  └────┬─────┘  └────┬─────┘      │
│       │              │              │            │
│       └──────────────┼──────────────┘            │
│                      │                           │
└──────────────────────┼───────────────────────────┘
                       │ MCP
               ┌───────┴───────┐
               │    Mimir      │
               │  (SQLite+FTS) │
               └───────────────┘

Requirements

• Python 3.11+
• Band.ai account (free Pro with code BANDHACK26)
• Mimir (runs locally, no cloud dependency)
• OpenAI API key (for agent LLM)

License

MIT — see LICENSE.

This site is open source. Improve this page.