Skip to content

CONFIG.md

Latest commit

 

History

History
327 lines (273 loc) · 10.2 KB

CONFIG.md

File metadata and controls

327 lines (273 loc) · 10.2 KB

Configuration

Configuration File

Default location: ~/.kimi/mneme/config.json

Created automatically on first run with sensible defaults.

Full Configuration Reference

{
  "db": {
    "path": "~/.kimi/mneme/mneme.db",
    "backup_enabled": true,
    "backup_interval_days": 7
  },
  "vector": {
    "path": "~/.kimi/mneme/vectors",
    "embedding_model": "sentence-transformers/all-MiniLM-L6-v2",
    "chunk_size": 512,
    "chunk_overlap": 50
  },
  "llm": {
    "provider": "kimi",
    "model": "kimi-k2.5",
    "base_url": null,
    "api_key": null,
    "timeout": 60.0,
    "options": {}
  },
  "compression": {
    "enabled": true,
    "provider": null,
    "model": null,
    "batch_size": 10,
    "min_observations": 5,
    "trigger": "session_end"
  },
  "structuring": {
    "enabled": true,
    "provider": null,
    "model": null,
    "fallback_to_heuristic": true,
    "heuristic_threshold_chars": 300,
    "batch_size": 5,
    "worker_interval_seconds": 5,
    "max_retry_count": 3
  },
  "injection": {
    "enabled": true,
    "max_tokens": 1500,
    "min_relevance": 0.7,
    "max_results": 2,
    "recency_boost_days": 7,
    "format": "markdown",
    "use_vector": false
  },
  "server": {
    "enabled": true,
    "auto_start": true,
    "host": "127.0.0.1",
    "port": 37777,
    "cors_origins": ["http://localhost:37777", "http://127.0.0.1:37777"],
    "loop": "auto"
  },
  "mcp": {
    "enabled": true,
    "auto_start": false,
    "transport": "stdio",
    "port": 37778
  },
  "privacy": {
    "exclude_tags": ["<private>", "<secret>"],
    "exclude_patterns": [
      "*.env*",
      "*.env.local",
      "*secret*",
      "*password*",
      "*token*",
      "*api_key*",
      "*private_key*"
    ],
    "max_content_length": 100000
  },
  "hooks": {
    "fire_and_forget": true,
    "timeout_seconds": 30,
    "batch_writes": true,
    "batch_interval_seconds": 5
  },
  "logging": {
    "level": "INFO",
    "file": "~/.kimi/mneme/mneme.log",
    "max_size_mb": 10,
    "backup_count": 5
  }
}

Configuration Sections

Database

Note: sqlite3 CLI is required for database inspection and internal operations. Install via your system package manager (apt install sqlite3, brew install sqlite3, winget install SQLite.SQLite, etc.).

Key Type Default Description
db.path string ~/.kimi/mneme/mneme.db SQLite database file path
db.backup_enabled boolean true Enable automatic backups
db.backup_interval_days integer 7 Days between backups

Vector Store

Key Type Default Description
vector.path string ~/.kimi/mneme/vectors Vector store directory (embeddings cache)
vector.embedding_model string sentence-transformers/all-MiniLM-L6-v2 Model for embeddings
vector.chunk_size integer 512 Text chunk size for embedding
vector.chunk_overlap integer 50 Overlap between chunks

LLM (Global Settings)

The llm section defines the default provider for all AI-powered features (structuring, compression). Individual sections (compression, structuring) can override these with their own provider and model.

Key Type Default Description
llm.provider string kimi LLM provider: kimi, ollama, openai_compatible
llm.model string kimi-k2.5 Model name (provider-specific)
llm.base_url string null Custom base URL for the API
llm.api_key string null API key or token
llm.timeout float 60.0 Request timeout in seconds
llm.options object {} Provider-specific extra options

Provider Examples

Kimi / Moonshot (default — tries OAuth first, falls back to API key):

When you log in to Kimi CLI via kimi login, mneme will automatically try to reuse that OAuth token. However, the Kimi Code API (api.kimi.com/coding/v1) is restricted to official Coding Agents and may return 403 Forbidden. If that happens, mneme falls back to heuristic structuring (rule-based, no AI).

To enable full AI structuring, provide a Moonshot API key from https://platform.moonshot.cn:

export MOONSHOT_API_KEY="sk-your-key-here"

Or configure explicitly in config.json:

{
  "llm": {
    "provider": "openai_compatible",
    "model": "moonshot-v1-8k",
    "base_url": "https://api.moonshot.cn/v1",
    "api_key": "sk-your-key-here"
  }
}

Ollama (local, free, no API key needed):

{
  "llm": {
    "provider": "ollama",
    "model": "llama3.2",
    "base_url": "http://localhost:11434"
  }
}

OpenAI-compatible (vLLM, LM Studio, etc.):

{
  "llm": {
    "provider": "openai_compatible",
    "model": "qwen2.5-coder",
    "base_url": "http://localhost:8000/v1",
    "api_key": "sk-optional"
  }
}

Compression (AI Summarization)

Key Type Default Description
compression.enabled boolean true Enable AI compression
compression.provider string null Override llm.provider (null = use global)
compression.model string null Override llm.model (null = use global)
compression.batch_size integer 10 Observations per batch
compression.min_observations integer 5 Min observations to compress
compression.trigger string session_end When to compress: session_end, threshold, manual

Structuring

Key Type Default Description
structuring.enabled boolean true Enable AI structuring
structuring.provider string null Override llm.provider (null = use global)
structuring.model string null Override llm.model (null = use global)
structuring.fallback_to_heuristic boolean true Use heuristic when AI fails
structuring.heuristic_threshold_chars integer 300 Use heuristic for short outputs
structuring.batch_size integer 5 Observations per batch
structuring.worker_interval_seconds integer 5 Worker poll interval
structuring.max_retry_count integer 3 Max retries per observation

Context Injection

Key Type Default Description
injection.enabled boolean true Inject past context at session start
injection.max_tokens integer 1500 Max tokens to inject
injection.min_relevance float 0.7 Minimum relevance score (0.0–1.0)
injection.max_results integer 2 Max sessions to inject
injection.recency_boost_days integer 7 Only consider sessions within N days
injection.format string markdown Output format: markdown, json, plain
injection.use_vector boolean false Use vector search for relevance scoring

Web Server

Key Type Default Description
server.enabled boolean true Start web server
server.auto_start boolean true Auto-start server on launch
server.host string 127.0.0.1 Bind address
server.port integer 37777 Port number
server.cors_origins array ["http://localhost:37777", "http://127.0.0.1:37777"] Allowed CORS origins

| server.loop | string | auto | Event loop: auto, asyncio, winloop (Windows only) |

MCP Server

Key Type Default Description
mcp.enabled boolean true Enable MCP server
mcp.auto_start boolean false Auto-start MCP with FastAPI
mcp.transport string stdio Transport: stdio or sse
mcp.port integer 37778 Port for SSE transport

Privacy

Key Type Default Description
privacy.exclude_tags array ["<private>", "<secret>"] HTML-like tags to exclude
privacy.exclude_patterns array ["*.env*", ...] Glob patterns to exclude
privacy.max_content_length integer 100000 Max characters to store

Hooks

Key Type Default Description
hooks.fire_and_forget boolean true Non-blocking hook execution
hooks.timeout_seconds integer 30 Hook timeout
hooks.batch_writes boolean true Batch DB writes
hooks.batch_interval_seconds integer 5 Batch flush interval

Logging

Key Type Default Description
logging.level string INFO Log level: DEBUG, INFO, WARNING, ERROR
logging.file string ~/.kimi/mneme/mneme.log Log file path
logging.max_size_mb integer 10 Max log file size
logging.backup_count integer 5 Number of backup files

Environment Variables

All config values can be overridden via environment variables:

# Database
export MNEME_DB_PATH="/custom/path/mneme.db"

# Vector store
# Vector store (embeddings cache, optional)
# export MNEME_VECTOR_PATH="/custom/path/vectors"

# LLM provider
export MNEME_LLM_PROVIDER="ollama"
export MNEME_LLM_MODEL="llama3.2"
export MNEME_LLM_BASE_URL="http://localhost:11434"
export MNEME_LLM_API_KEY="sk-..."

# Moonshot API key (alternative to Kimi OAuth)
export MOONSHOT_API_KEY="sk-..."

# Structuring
export MNEME_STRUCTURING_ENABLED="true"

# Server
export MNEME_SERVER_PORT="37777"
export MNEME_SERVER_HOST="0.0.0.0"

# Logging
export MNEME_LOG_LEVEL="DEBUG"

Environment variables take precedence over config file values.

Per-Project Configuration

Create .mneme.json in project root for project-specific settings:

{
  "llm": {
    "provider": "ollama",
    "model": "codellama"
  },
  "privacy": {
    "exclude_patterns": ["*.local.env", "secrets/"]
  },
  "injection": {
    "max_tokens": 1000,
    "recency_boost_days": 14,
    "include_patterns": true
  },
  "compression": {
    "min_observations": 3
  }
}

Project config merges with global config (project values override global). The file is searched in:

  1. Current working directory
  2. Parent directories (up to filesystem root)

This makes it work automatically when you cd into a project.