Configuration

mindroom config validate --path /path/to/config.yaml

# Agent definitions (at least one recommended)
agents:
  assistant:
    display_name: Assistant        # Required: Human-readable name
    role: A helpful AI assistant   # Optional: Description of purpose
    model: sonnet                  # Optional: Model name (default: "default")
    tools:                         # Optional: Agent-specific tools (merged with defaults.tools)
      - file
      - shell
      # - shell:                    # Per-agent tool config overrides (single-key dict):
      #     extra_env_passthrough: "DAWARICH_*"
    include_default_tools: true    # Optional: Per-agent opt-out for defaults.tools
    skills: []                     # Optional: List of skill names
    instructions: []               # Optional: Custom instructions
    rooms: [lobby]                 # Optional: Rooms to auto-join
    markdown: true                 # Optional: Override default (inherits from defaults section)
    worker_tools: [shell, file]    # Optional: Override default (inherits from defaults section)
    worker_scope: user_agent       # Optional: Reuse one proxied runtime per requester+agent
    learning: true                 # Optional: Override default (inherits from defaults section)
    learning_mode: always          # Optional: Override default (inherits from defaults section)
    memory_backend: file           # Optional: Per-agent memory backend override (mem0 or file)
    knowledge_bases: [docs]         # Optional: Assign one or more configured knowledge bases
    context_files:                 # Optional: Load files into each freshly built agent instance
      - SOUL.md
      - AGENTS.md
      - USER.md
      - IDENTITY.md
      - TOOLS.md
      - HEARTBEAT.md
  researcher:
    display_name: Researcher
    role: Research and gather information
    model: sonnet
  writer:
    display_name: Writer
    role: Write and edit content
    model: sonnet
  developer:
    display_name: Developer
    role: Write code and implement features
    model: sonnet
  reviewer:
    display_name: Reviewer
    role: Review code and provide feedback
    model: sonnet

# Model configurations (at least a "default" model is recommended)
models:
  default:
    provider: anthropic            # Required: openai, anthropic, ollama, google, gemini, vertexai_claude, groq, cerebras, openrouter, deepseek
    id: claude-sonnet-4-6            # Required: Model ID for the provider
  sonnet:
    provider: anthropic            # Required: openai, anthropic, ollama, google, gemini, vertexai_claude, groq, cerebras, openrouter, deepseek
    id: claude-sonnet-4-6            # Required: Model ID for the provider
    host: null                     # Optional: Host URL (e.g., for Ollama)
    api_key: null                  # Optional: API key (usually from env vars)
    extra_kwargs: null             # Optional: Provider-specific parameters
    context_window: null           # Optional: Needed on the active runtime model for replay safety; explicit compaction.model also needs its own window for summary generation

# Team configurations (optional)
teams:
  research_team:
    display_name: Research Team    # Required: Human-readable name
    role: Collaborative research   # Required: Description of team purpose
    agents: [researcher, writer]   # Required: List of agent names
    mode: collaborate              # Optional: "coordinate" or "collaborate" (default: coordinate)
    model: sonnet                  # Optional: Model for team coordination (default: "default")
    num_history_runs: 8            # Optional: Team-scoped replay policy
    num_history_messages: null     # Optional: Mutually exclusive with num_history_runs
    max_tool_calls_from_history: 6 # Optional: Limit replayed tool call messages
    compaction:                    # Optional: Team-scoped auto-compaction overrides
      enabled: true
      threshold_percent: 0.8
      reserve_tokens: 16384
      notify: false
    rooms: []                      # Optional: Rooms to auto-join

# Culture configurations (optional)
cultures:
  engineering:
    description: Follow clean code principles and write tests  # Shared principles
    agents: [developer, reviewer]  # Agents assigned (each agent can belong to at most one culture)
    mode: automatic                # automatic, agentic, or manual

# Router configuration (optional)
router:
  model: default                   # Optional: Model for routing (default: "default")

# Default settings for all agents (optional)
defaults:
  tools:                           # Default: ["scheduler"] (added to every agent; set [] to disable)
    - scheduler                    # Plain string or single-key dict with inline config overrides
  markdown: true                   # Default: true
  enable_streaming: true           # Default: true (stream responses via message edits)
  streaming:
    update_interval: 5.0           # Default: 5.0 (steady-state seconds between streamed edits)
    min_update_interval: 0.5       # Default: 0.5 (fast-start seconds between early edits)
    interval_ramp_seconds: 15.0    # Default: 15.0 (set 0 to disable interval ramping)
  learning: true                   # Default: true
  learning_mode: always            # Default: always (or agentic)
  max_preload_chars: 50000         # Hard cap for preloaded context from context_files
  show_stop_button: true           # Default: true (global only, cannot be overridden per-agent)
  num_history_runs: null           # Number of prior runs to include (null = all)
  num_history_messages: null       # Max messages from history (null = use num_history_runs)
  compress_tool_results: true      # Compress tool results in history to save context
  # Auto-compaction stays off until you author defaults.compaction
  # or a non-empty per-agent/per-team compaction override.
  # A bare compaction: {} only inherits authored defaults.
  # compaction:
  #   enabled: true
  #   threshold_percent: 0.8
  #   reserve_tokens: 16384
  #   notify: false
  max_tool_calls_from_history: null  # Limit tool call messages replayed from history (null = no limit)
  show_tool_calls: true            # Default: true (show tool call details inline in responses)
  worker_tools: null               # Default: null (tool names to route through workers; null = use MindRoom's default routing policy, [] = disable)
  worker_scope: null               # Default: null (no runtime reuse; set shared/user/user_agent to enable)
  allow_self_config: false         # Default: false (allow agents to modify their own config via a tool)

# defaults.tools are appended to each agent's tools list with duplicates removed.
# Set agents.<name>.include_default_tools: false to opt out a specific agent.
# defaults.streaming is also global-only and controls streamed message edit cadence.
# Tools can be plain strings or single-key dicts with per-agent config overrides.

# Auto-compaction is destructive inside the active session.
# It rewrites the stored session summary and removes compacted raw runs from
# the live session so Agno replays only the summary plus recent runs.
# Use __MINDROOM_INHERIT__ inside a tool override to clear one inherited authored field
# while keeping the rest of defaults.tools for that agent.
# See agents.md for the full per-agent tool configuration syntax.

# Memory system configuration (optional)
memory:
  backend: mem0                    # Global default backend (mem0 or file); agents can override with memory_backend
  team_reads_member_memory: false  # Default: false (when true, team reads can access member agent memories)
  embedder:
    provider: openai               # Default: openai (openai, ollama, huggingface, sentence_transformers)
    config:
      model: text-embedding-3-small  # Default embedding model
      api_key: null                # Optional: From env var
      host: null                   # Optional: For self-hosted
      dimensions: null             # Optional: Embedding dimension override (e.g., 256)
  llm:                             # Optional: LLM for memory operations
    provider: ollama
    config: {}
  file:                            # File-backed memory settings (when backend: file)
    path: null                     # Optional: fallback root for file memory paths
    max_entrypoint_lines: 200      # Default: 200 (max lines preloaded from MEMORY.md)
  auto_flush:                      # Background memory auto-flush (file backend only)
    enabled: false                 # Default: false (enable background flush worker)
    flush_interval_seconds: 1800   # Default: 1800 (loop interval)
    idle_seconds: 120              # Default: 120 (idle time before flush eligibility)
    max_dirty_age_seconds: 600     # Default: 600 (force flush after this many seconds dirty)
    stale_ttl_seconds: 86400       # Default: 86400 (drop stale flush-state entries older than this)
    max_cross_session_reprioritize: 5  # Default: 5 (same-agent dirty sessions reprioritized per prompt)
    retry_cooldown_seconds: 30     # Default: 30 (cooldown before retrying a failed extraction)
    max_retry_cooldown_seconds: 300  # Default: 300 (upper bound for retry cooldown backoff)
    batch:
      max_sessions_per_cycle: 10   # Default: 10 (max sessions processed per auto-flush loop)
      max_sessions_per_agent_per_cycle: 3  # Default: 3 (max sessions per agent per loop)
    extractor:
      no_reply_token: NO_REPLY     # Default: NO_REPLY (token indicating no durable memory)
      max_messages_per_flush: 20   # Default: 20 (max messages considered per extraction)
      max_chars_per_flush: 12000   # Default: 12000 (max chars considered per extraction)
      max_extraction_seconds: 30   # Default: 30 (timeout for one extraction job)
      include_memory_context:
        memory_snippets: 5         # Default: 5 (max MEMORY.md snippets for dedupe context)
        snippet_max_chars: 400     # Default: 400 (max chars per snippet)
#
# See docs/memory.md for full auto-flush behavior and tuning guidance.
#
# Set memory.embedder.provider: sentence_transformers to run embeddings in-process.
# MindRoom auto-installs that optional extra on first use.

# Knowledge base configuration (optional)
knowledge_bases:
  docs:
    path: ./knowledge_docs          # Folder containing documents for this base (Pydantic default)
    watch: true                    # Reindex automatically when files change
    chunk_size: 5000               # Default: 5000 (max characters per indexed chunk)
    chunk_overlap: 0               # Default: 0 (overlapping characters between chunks)
    git:                           # Optional: Sync this folder from a Git repository
      repo_url: https://github.com/pipefunc/pipefunc
      branch: main
      poll_interval_seconds: 300
      skip_hidden: true
      include_patterns: ["docs/**"]  # Optional: root-anchored glob filters
      exclude_patterns: []
      credentials_service: github_private # Optional: service in CredentialsManager

# Voice message handling (optional)
voice:
  enabled: false                   # Default: false
  visible_router_echo: false       # Optional: show the normalized voice text from the router
  stt:
    provider: openai               # Default: openai
    model: whisper-1               # Default: whisper-1
    api_key: null
    host: null
  intelligence:
    model: default                 # Model for command recognition

# Internal MindRoom user account (optional, omit for hosted/public profiles)
# When present, defaults are: username: mindroom_user, display_name: MindRoomUser
mindroom_user:
  username: mindroom_user          # Set before first startup (localpart only)
  display_name: MindRoomUser       # Can be changed later

# Matrix room onboarding/discoverability (optional)
matrix_room_access:
  mode: single_user_private        # Default keeps invite-only/private behavior
  multi_user_join_rule: public     # In multi_user mode: public or knock
  publish_to_room_directory: false # Publish managed rooms in server room directory
  invite_only_rooms: []            # Room keys/aliases/IDs that stay invite-only/private
  reconcile_existing_rooms: false  # Explicit migration of existing managed rooms

# Authorization (optional)
authorization:
  global_users: []                 # Users with access to all rooms
  room_permissions: {}             # Keys: room ID (!id), full alias (#alias:domain), or managed room key (alias)
  default_room_access: false       # Default: false
  aliases: {}                      # Map canonical Matrix user IDs to bridge aliases (see authorization docs)
  agent_reply_permissions: {}      # Per-agent/team/router (or '*') reply allowlists; supports globs like '*:example.com'

# Room-specific model overrides (optional)
# Keys are room aliases, values are model names from the models section
# Example: room_models: {dev: sonnet, lobby: gpt4o}
room_models: {}

# Non-MindRoom bot accounts to exclude from multi-human detection (optional)
# These accounts won't trigger the mention requirement in threads
bot_accounts:
  - "@telegram:example.com"

# Plugin paths (optional)
plugins: []

# Matrix Space grouping (optional)
matrix_space:
  enabled: true                    # Default: true (create a root Matrix Space for managed rooms)
  name: MindRoom                   # Default: "MindRoom" (display name for the root Space)

# Timezone for scheduled tasks (optional)
timezone: America/Los_Angeles      # Default: UTC