Memoryblock uses a clean, hierarchical configuration system. Higher-level settings cascade down to blocks, with blocks able to override any inherited value.

Config Hierarchy

~/.memoryblock/              ← Global defaults (machine-wide)
├── config.json              ← Global config
└── auth.json                ← Global credentials (fallback)

~/.memoryblock/ws/           ← Your workspace (portable)
├── config.json              ← Workspace config (overrides global)
├── auth.json                ← Workspace credentials (priority)
├── founder.md               ← Founder profile
└── blocks/
    └── <block-name>/
        ├── config.json      ← Block config (overrides workspace defaults)
        ├── monitor.md       ← Monitor identity
        ├── memory.md        ← Persistent context
        ├── session.json     ← Active session state (auto-managed)
        ├── pulse.json       ← Autonomic pulse instructions & status
        ├── costs.json       ← Token/cost tracking
        └── logs/            ← Full conversation logs

Config priority: Block → Workspace defaults → Global ~/.memoryblock/

Auth priority: Workspace auth.json → Global ~/.memoryblock/auth.json

Workspace credentials are portable — move the folder anywhere and everything follows.

Every block has a config.json that controls its behavior. Here’s a complete reference.

Adapter Settings

{
    "adapter": {
        "provider": "bedrock",
        "model": "us.anthropic.claude-sonnet-4-5-20250929-v1:0",
        "region": "us-east-1",
        "maxTokens": 4096,
        "cacheControl": false,
        "baseURL": null
    }
}

Where the provider list comes from

memoryblock no longer ships per-provider adapter code. The provider list comes from the bundled aiplug runtime — every provider in the world (OpenAI, Anthropic, Bedrock, Gemini, Ollama, MiniMax, DeepSeek, Moonshot, GLM, Zhipu, Together, Groq, vLLM, llama.cpp server, …) is sourced from aiplug’s registry. Adding a new provider is an aiplug change, not a memoryblock change.

The provider registry page has the full list and the discovery workflow.

auth.json — Provider Credentials

The auth.json file in your workspace holds credentials keyed by provider. Each provider section is optional — only the ones you use need to be filled in.

{
  "providers": {
    "bedrock":   { "apiKey": "AKIA:secret", "region": "us-east-1" },
    "openai":    { "apiKey": "sk-proj-..." },
    "anthropic": { "apiKey": "sk-ant-..." },
    "minimax":   { "apiKey": "..." }
  },
  "telegram":  { "botToken": "8445465...", "chatId": "5315436002" },
  "brave":     { "apiKey": "BSA..." }
}

The new providers object is the canonical place. Each key is an aiplug provider slug; each value is whatever credentials that provider needs (the schema is discovered from AIPlug.configSchema(slug) at runtime). The legacy flat fields — aws, openai, google, anthropic, gemini, ollama — are still honoured and merged into the new shape by loadAuth(), so existing auth.json files keep working without migration.

Legacy gemini key. Blocks written before the rename to google will still find their API key under the old gemini key — it’s auto-mirrored at load time. New code should use google.

Ollama Setup State

When you pick Ollama during mblk init, memoryblock records install + model state at ~/.memoryblock/ws/ollama.json:

{
  "installed": true,
  "daemonManaged": true,
  "defaultModel": "qwen2.5:0.5b",
  "installedAt": "2026-06-12T19:50:00.000Z",
  "lastChecked": "2026-06-12T20:05:23.142Z"
}

Re-running mblk init is silent once this file is present. Delete it to re-trigger the install/serve/pull prompts.

Tools Configuration

{
    "tools": {
        "enabled": [
            "read_file", "write_file", "list_directory", "create_directory",
            "execute_command", "web_search",
            "search_files", "replace_in_file", "file_info",
            "run_lint", "run_build", "run_test"
        ],
        "searchProvider": "brave",
        "sandbox": true,
        "workingDir": "/path/to/project"
    }
}

Available Tools

Safe commands that auto-execute without approval: ls, cat, grep, find, git status/log/diff, pnpm lint, pnpm build, pnpm test, etc.

Memory Settings

{
    "memory": {
        "maxContextTokens": 100000,
        "thresholdPercent": 80
    }
}

Execution Permissions

Blocks are structurally restricted across bash commands and network fetch tooling using explicit security parameters visible directly within the Dashboard layout:

{
    "permissions": {
        "scope": "block",
        "allowShell": false,
        "allowNetwork": true,
        "maxTimeout": 120000
    }
}

Monitor Identity

The monitor sets these during its first onboarding conversation:

{
    "monitorName": "Sam",
    "monitorEmoji": "🌟"
}

These persist across sessions and appear in message headers: :: home / 🌟 Sam

Block Files

Every block contains:

blocks/my-block/
├── config.json     # Block configuration
├── pulse.json      # Autonomic instructions & status
├── monitor.md      # Monitor identity, personality, notes
├── memory.md       # Smart context log (auto-managed)
├── session.json    # Active session state (for resumption)
├── chat.json       # WebChannel interactive queue payload
├── costs.json      # Token usage and cost tracking
├── logs/           # Conversation logs (timestamped .txt)
└── agents/         # Agent definitions (if using agent plugin)

Cost Tracking

Costs are tracked system-level (zero model tokens wasted). After each response, the session cost is displayed. Cumulative costs persist in costs.json.

Example: Developer Block (Opus 4.6)

{
    "name": "builder",
    "description": "Self-development block with full capabilities",
    "adapter": {
        "provider": "bedrock",
        "model": "us.anthropic.claude-opus-4-6-v1",
        "maxTokens": 8192,
        "cacheControl": false
    },
    "tools": {
        "sandbox": false,
        "workingDir": "/path/to/project"
    },
    "memory": {
        "maxContextTokens": 200000,
        "thresholdPercent": 80
    }
}

Start with: mblk start builder (CLI) or mblk start builder --channel telegram