# MemOS Environment Variables (core runtime)
# Legend: [required] needed for default startup; others are optional or conditional per comments.

## Base
TZ=Asia/Shanghai
MOS_CUBE_PATH=/tmp/data_test               # local data path
MEMOS_BASE_PATH=.                          # CLI/SDK cache path
MOS_ENABLE_DEFAULT_CUBE_CONFIG=true        # enable default cube config
MOS_ENABLE_REORGANIZE=false                # enable memory reorg
# MOS Text Memory Type
MOS_TEXT_MEM_TYPE=general_text             # general_text | tree_text
ASYNC_MODE=sync                            # async/sync, used in default cube config

## User/session defaults
# Top-K for LLM in the Product API（old version）
MOS_TOP_K=50

## Chat LLM (main dialogue)
# LLM model name for the Product API
MOS_CHAT_MODEL=gpt-4o-mini
# Temperature for LLM in the Product API
MOS_CHAT_TEMPERATURE=0.8
# Max tokens for LLM in the Product API
MOS_MAX_TOKENS=2048
# Top-P for LLM in the Product API
MOS_TOP_P=0.9
# LLM for the Product API backend
MOS_CHAT_MODEL_PROVIDER=openai             # openai | huggingface | vllm | minimax
OPENAI_API_KEY=sk-xxx                      # [required] when provider=openai
OPENAI_API_BASE=https://api.openai.com/v1  # [required] base for the key
# Shared Qwen credentials for task-specific LLMs selected by qwen* model names
QWEN_API_KEY=                              # e.g. Bailian/DashScope API key
QWEN_API_BASE=https://dashscope.aliyuncs.com/compatible-mode/v1
# MiniMax LLM (when provider=minimax)
# MINIMAX_API_KEY=your-minimax-api-key     # [required] when provider=minimax
# MINIMAX_API_BASE=https://api.minimax.io/v1  # base for MiniMax API

## MemReader / retrieval LLM
MEMRADER_MODEL=gpt-4o-mini
MEMRADER_API_KEY=sk-xxx                    # [required] can reuse OPENAI_API_KEY
MEMRADER_API_BASE=http://localhost:3000/v1 # [required] base for the key
MEMRADER_MAX_TOKENS=5000
MEMREADER_ENABLE_BACKUP=false              # fallback from MEMRADER_MODEL to MEMREADER_GENERAL_MODEL on primary failure

## Task-specific LLMs
# These settings only specify model names. API key/base are selected from QWEN_* or OPENAI_* based on the model family.
MEMREADER_GENERAL_MODEL=                   # used by hallucination filter, rewrite, merge, skill extraction, scheduler process_llm
PREFERENCE_EXTRACTOR_MODEL=                # falls back to MEMREADER_GENERAL_MODEL when omitted
FEEDBACK_MODEL=                            # falls back to MEMREADER_GENERAL_MODEL when omitted
SUGGESTION_MODEL=                          # falls back to MEMREADER_GENERAL_MODEL when omitted
MEMSCHEDULER_MODEL=                        # falls back to MEMREADER_GENERAL_MODEL when omitted
DOCUMENT_PARSER_MODEL=                     # falls back to MEMREADER_GENERAL_MODEL when omitted
IMAGE_PARSER_MODEL=                        # falls back to MEMREADER_GENERAL_MODEL when omitted
QWEN_MODEL=qwen-flash                      # optional qwen_llm slot when QWEN_API_KEY is set

## Optional per-model LLM QPS rate limiting (Redis GCRA)
# Disabled by default. Enable explicitly and tune rules to your provider quota.
MEMOS_LLM_RATE_LIMIT_ENABLED=false
MEMOS_LLM_RATE_LIMIT_RULES='{"gpt-4o-mini":{"qps":5,"burst":2,"max_wait_seconds":30,"queue_capacity":16,"retry_attempts":1}}'
# Rule keys must match actual model names; unlisted models are not limited.
# QPS/burst are shared across workers using the same Redis/DB and model key.
# queue_capacity is per process/model; max_wait_seconds is the permit-wait budget.
# Reuses MEMSCHEDULER_REDIS_HOST/PORT/DB/USERNAME/PASSWORD/SSL; host is required when enabled.
# See docs/cn/open_source/open_source_api/help/llm_qps_rate_limit.md.

## Embedding & rerank
# embedding dim
EMBEDDING_DIMENSION=1024
# set default embedding backend
MOS_EMBEDDER_BACKEND=universal_api        # universal_api | ollama
# set openai style
MOS_EMBEDDER_PROVIDER=openai              # required when universal_api
# embedding model
MOS_EMBEDDER_MODEL=bge-m3                 # siliconflow → use BAAI/bge-m3
# embedding url
MOS_EMBEDDER_API_BASE=http://localhost:8000/v1 # required when universal_api
# embedding model key
MOS_EMBEDDER_API_KEY=EMPTY                 # required when universal_api
MOS_EMBEDDER_HEADERS_EXTRA={}              # optional JSON headers for universal_api
MOS_EMBEDDER_BACKUP_CLIENT=false
MOS_EMBEDDER_BACKUP_API_BASE=http://openai.com
MOS_EMBEDDER_BACKUP_API_KEY=sk-xxxx
MOS_EMBEDDER_BACKUP_HEADERS_EXTRA={}
MOS_EMBEDDER_BACKUP_MODEL=text-embedding-3-large
OLLAMA_API_BASE=http://localhost:11434     # required when backend=ollama
# reranker config
MOS_RERANKER_BACKEND=http_bge             # http_bge | http_bge_strategy | cosine_local
# reranker url
MOS_RERANKER_URL=http://localhost:8001     # required when backend=http_bge*
# reranker model
MOS_RERANKER_MODEL=bge-reranker-v2-m3     # siliconflow → use BAAI/bge-reranker-v2-m3
MOS_RERANKER_HEADERS_EXTRA={}             # extra headers, JSON string, e.g. {"Authorization":"Bearer your_token"}
# use source
MOS_RERANKER_STRATEGY=single_turn
MOS_FEEDBACK_RERANKER_BACKEND=http_bge
MOS_FEEDBACK_RERANKER_MODEL=bge-reranker-v2-m3
MOS_RERANKER_MAX_TOKENS=8000
MOS_RERANKER_CONCAT_LEN=1000


# External Services (for evaluation scripts)
# API key for reproducting Zep(compertitor product)  evaluation
ZEP_API_KEY=your_zep_api_key_here
# API key for reproducting Mem0(compertitor product)  evaluation
MEM0_API_KEY=your_mem0_api_key_here
# API key for reproducting MemU(compertitor product)  evaluation
MEMU_API_KEY=your_memu_api_key_here
# API key for reproducting MEMOBASE(compertitor product)  evaluation
MEMOBASE_API_KEY=your_memobase_api_key_here
# Project url for reproducting MEMOBASE(compertitor product)  evaluation
MEMOBASE_PROJECT_URL=your_memobase_project_url_here
# LLM for evaluation
MODEL=gpt-4o-mini
# embedding model for evaluation
EMBEDDING_MODEL=nomic-embed-text:latest


## Internet search & preference memory
# Enable web search
ENABLE_INTERNET=false
# Internet search backend (bocha | tavily)
INTERNET_SEARCH_BACKEND=bocha
# API key for BOCHA Search
BOCHA_API_KEY=                             your-bocha-api-key and backend=bocha
# API key for Tavily Search
TAVILY_API_KEY=                            your-bocha-api-key and backend=tavily
# default search mode
SEARCH_MODE=fast                          # fast | fine | mixture
# Slow retrieval strategy configuration, rewrite is the rewrite strategy
FINE_STRATEGY=rewrite                     # rewrite | recreate | deep_search
# Whether to enable preference memory
ENABLE_PREFERENCE_MEMORY=true
# Preference Memory Add Mode
PREFERENCE_ADDER_MODE=fast                # fast | safe
# Whether to deduplicate explicit preferences based on factual memory
DEDUP_PREF_EXP_BY_TEXTUAL=false
PROFILE_MEMORY_RESERVED_TOP_K=0

## Reader chunking
MEM_READER_BACKEND=simple_struct           # simple_struct | strategy_struct
MEM_READER_CHAT_CHUNK_TYPE=default         # default | content_length
MEM_READER_CHAT_CHUNK_TOKEN_SIZE=1600      # tokens per chunk (default mode)
MEM_READER_CHAT_CHUNK_SESS_SIZE=10         # sessions per chunk (default mode)
MEM_READER_CHAT_CHUNK_OVERLAP=2            # overlap between chunks
MEM_READER_SAVE_RAWFILENODE=true
MEM_READER_MEM_VERSION_SWITCH=off
FILE_PARSER_DIRECT_MARKDOWN_HOSTNAMES=139.196.232.20
SKILLS_OSS_DIR=skill_memory/
SKILLS_LOCAL_TMP_DIR=/tmp/skill_memory/
SKILLS_LOCAL_DIR=/tmp/upload_skill_memory/

## Scheduler (MemScheduler / API)
# Enable or disable the main switch for configuring the memory scheduler during MemOS class initialization
MOS_ENABLE_SCHEDULER=false
# Determine the number of most relevant memory entries that the scheduler retrieves or processes during runtime (such as reordering or updating working memory)
MOS_SCHEDULER_TOP_K=10
# The time interval (in seconds) for updating "Activation Memory" (usually referring to caching or short-term memory mechanisms)
MOS_SCHEDULER_ACT_MEM_UPDATE_INTERVAL=300
# The size of the context window considered by the scheduler when processing tasks (such as the number of recent messages or conversation rounds)
MOS_SCHEDULER_CONTEXT_WINDOW_SIZE=5
# The maximum number of working threads allowed in the scheduler thread pool for concurrent task execution
MOS_SCHEDULER_THREAD_POOL_MAX_WORKERS=10000
# The polling interval (in seconds) for the scheduler to consume new messages/tasks from the queue. The smaller the value, the faster the response, but the CPU usage may be higher
MOS_SCHEDULER_CONSUME_INTERVAL_SECONDS=0.01
# Whether to enable the parallel distribution function of the scheduler to improve the throughput of concurrent operations
MOS_SCHEDULER_ENABLE_PARALLEL_DISPATCH=true
# The specific switch to enable or disable the "Activate Memory" function in the scheduler logic
MOS_SCHEDULER_ENABLE_ACTIVATION_MEMORY=false
# Control whether the scheduler instance is actually started during server initialization. If false, the scheduler object may be created but its background loop will not be started
API_SCHEDULER_ON=true
# Specifically define the window size for API search operations in OptimizedScheduler. It is passed to the ScherderrAPIModule to control the scope of the search context
API_SEARCH_WINDOW_SIZE=5
# Specify how many rounds of previous conversations (history) to retrieve and consider during the 'hybrid search' (fast search+asynchronous fine search). This helps provide context aware search results
API_SEARCH_HISTORY_TURNS=5
MEMSCHEDULER_USE_REDIS_QUEUE=false
MEMSCHEDULER_STREAM_KEY_PREFIX=
MEMSCHEDULER_REDIS_STREAM_KEY_PREFIX=
MEMSCHEDULER_REDIS_REFILL_TIMEOUT_SEC=30
MEMSCHEDULER_REDIS_CACHE_MAX_PACKS=50
MEMSCHEDULER_REDIS_INITIAL_SCAN_MAX_KEYS=1000
MEMSCHEDULER_REDIS_INITIAL_SCAN_TIME_LIMIT_SEC=1.0
MEMSCHEDULER_REDIS_PIPELINE_CHUNK_SIZE=200

## Graph / vector stores
# Neo4j database selection mode
NEO4J_BACKEND=neo4j-community             # neo4j-community | neo4j | polardb | postgres
# Neo4j database url
NEO4J_URI=bolt://localhost:7687           # required when backend=neo4j*
#  Neo4j database user
NEO4J_USER=neo4j                          # required when backend=neo4j*
# Neo4j database password
NEO4J_PASSWORD=12345678                   # required when backend=neo4j*
# Neo4j database name
NEO4J_DB_NAME=neo4j                       # required for shared-db mode
# Neo4j database data sharing with Memos
MOS_NEO4J_SHARED_DB=false
QDRANT_HOST=localhost
QDRANT_PORT=6333
# For Qdrant Cloud / remote endpoint (takes priority if set):
QDRANT_URL=your_qdrant_url
QDRANT_API_KEY=your_qdrant_key
# milvus server uri
MILVUS_URI=http://localhost:19530         # required when ENABLE_PREFERENCE_MEMORY=true
MILVUS_USER_NAME=root                     # same as above
MILVUS_PASSWORD=12345678                  # same as above

# PolarDB endpoint/host
POLAR_DB_HOST=localhost
# PolarDB port
POLAR_DB_PORT=5432
# PolarDB username
POLAR_DB_USER=root
# PolarDB password
POLAR_DB_PASSWORD=123456
# PolarDB database name
POLAR_DB_DB_NAME=shared_memos_db
# PolarDB Server Mode:
# If set to true, use Multi-Database Mode where each user has their own independent database (physical isolation).
# If set to false (default), use Shared Database Mode where all users share one database with logical isolation via username.
POLAR_DB_USE_MULTI_DB=false
# PolarDB connection pool size
POLARDB_POOL_MAX_CONN=100
CONNECTION_WAIT_TIMEOUT=60
SKIP_CONNECTION_HEALTH_CHECK=false
WARM_UP_ON_STARTUP_BY_FULL=false
WARM_UP_ON_STARTUP_BY_ALL=false

## PostgreSQL / pgvector graph backend
POSTGRES_HOST=postgres
POSTGRES_PORT=5432
POSTGRES_USER=n8n
POSTGRES_PASSWORD=
POSTGRES_DB=n8n
MEMOS_SCHEMA=memos
POSTGRES_MAX_CONN=20

## MySQL user manager
MOS_USER_MANAGER_BACKEND=sqlite            # sqlite | mysql
MYSQL_HOST=localhost
MYSQL_PORT=3306
MYSQL_USERNAME=root
MYSQL_PASSWORD=12345678
MYSQL_DATABASE=memos_users
MYSQL_CHARSET=utf8mb4

## Related configurations of Redis
# Reddimq sends scheduling information and synchronization information for some variables
MEMSCHEDULER_REDIS_HOST=                  # fallback keys if not using the global ones
MEMSCHEDULER_REDIS_PORT=
MEMSCHEDULER_REDIS_DB=
MEMSCHEDULER_REDIS_PASSWORD=
MEMSCHEDULER_REDIS_TIMEOUT=
MEMSCHEDULER_REDIS_CONNECT_TIMEOUT=


## Nacos (optional config center)
# Nacos turns off long polling listening, defaults to true
NACOS_ENABLE_WATCH=false
# The monitoring interval for long rotation training is 60 seconds, and the default 30 seconds can be left unconfigured
NACOS_WATCH_INTERVAL=60
# nacos server address
NACOS_SERVER_ADDR=
# nacos dataid
NACOS_DATA_ID=
# nacos group
NACOS_GROUP=DEFAULT_GROUP
# nacos namespace
NACOS_NAMESPACE=
# nacos ak
AK=
# nacos sk
SK=

# chat model for chat api
CHAT_MODEL_LIST='[{
  "backend": "deepseek",
  "api_base": "http://localhost:1234",
  "api_key": "your-api-key",
  "model_name_or_path": "deepseek-r1",
  "support_models": ["deepseek-r1"]
}]'

# RabbitMQ host name for message-log pipeline
MEMSCHEDULER_RABBITMQ_HOST_NAME=
# RabbitMQ user name for message-log pipeline
MEMSCHEDULER_RABBITMQ_USER_NAME=
# RabbitMQ password for message-log pipeline
MEMSCHEDULER_RABBITMQ_PASSWORD=
# RabbitMQ virtual host for message-log pipeline
MEMSCHEDULER_RABBITMQ_VIRTUAL_HOST=memos
# Erase connection state on connect for message-log pipeline
MEMSCHEDULER_RABBITMQ_ERASE_ON_CONNECT=true
# RabbitMQ port for message-log pipeline
MEMSCHEDULER_RABBITMQ_PORT=5672
MEMSCHEDULER_RABBITMQ_EXCHANGE_NAME=
MEMSCHEDULER_RABBITMQ_EXCHANGE_TYPE=
MEMSCHEDULER_RABBITMQ_PLAYGROUND_CHAT_EXCHANGE_NAME=
MEMSCHEDULER_RABBITMQ_PLAYGROUND_CHAT_EXCHANGE_TYPE=

# Optional model list for /chat/stream/playground. Falls back to CHAT_MODEL_LIST when omitted.
# PLAYGROUND_CHAT_MODEL_LIST='[{"backend":"qwen","api_base":"https://dashscope.aliyuncs.com/compatible-mode/v1","api_key":"your-api-key","model_name_or_path":"qwen3-max-preview","support_models":["qwen3-max-preview"]}]'