Configuration

All configuration is managed through environment variables or a .env file. The install.sh script auto-generates it.

Setup

cp .env.example .env
vim .env

Environment Variables

Variable	Default	Description
AWS_REGION	us-east-1	AWS region
VECTOR_STORE	opensearch	Vector engine: opensearch or s3vectors
OPENSEARCH_HOST	localhost	OpenSearch host
OPENSEARCH_PORT	9200	OpenSearch port
OPENSEARCH_USER	admin	Username
OPENSEARCH_PASSWORD	—	Password
OPENSEARCH_USE_SSL	false	Enable SSL
OPENSEARCH_VERIFY_CERTS	false	Verify SSL certificates
OPENSEARCH_COLLECTION	mem0_memories	Index name
S3VECTORS_BUCKET_NAME	—	S3Vectors bucket (required for s3vectors mode)
S3VECTORS_INDEX_NAME	mem0	S3Vectors index name
EMBEDDING_MODEL	amazon.titan-embed-text-v2:0	Embedding model
EMBEDDING_DIMS	1024	Vector dimensions
LLM_MODEL	minimax.minimax-m2.5	LLM model for mem0 memory extraction / dedup (see LLM selection guide)
LLM_TEMPERATURE	0.1	LLM temperature
LLM_MAX_TOKENS	2000	Max tokens
DIGEST_LLM_MODEL	minimax.minimax-m2.5	LLM model for auto_digest.py summarization pipeline (defaults to LLM_MODEL if unset)
DIGEST_LLM_REGION	same as AWS_REGION	AWS region override for the digest LLM
SERVICE_HOST	0.0.0.0	Service bind address
SERVICE_PORT	8230	Service port
MEM0_TELEMETRY	true	⚠️ 建议关闭。 mem0 默认开启匿名遥测，每次 add/search/delete 都向 us.i.posthog.com 上报操作元数据（collection 名、LLM 类型、vector store 类型等），同时每次调用都创建一个新的 PostHog Consumer 线程，长时间运行后（尤其是 auto_dream 批量处理记忆后）会积累大量僵尸线程（测试中观察到 135 个）。私有部署请设为 false。
SEARCH_TOP_K	5	Default top_k for /memory/search and /memory/search_combined
SEARCH_RECENT_DAYS	3	Default recent days window for /memory/search_combined

Example .env

AWS_REGION=us-east-1

VECTOR_STORE=opensearch
OPENSEARCH_HOST=your-opensearch-host.es.amazonaws.com
OPENSEARCH_PORT=443
OPENSEARCH_USER=admin
OPENSEARCH_PASSWORD=your-password
OPENSEARCH_USE_SSL=true
OPENSEARCH_VERIFY_CERTS=true
OPENSEARCH_COLLECTION=mem0_memories

EMBEDDING_MODEL=amazon.titan-embed-text-v2:0
EMBEDDING_DIMS=1024

# LLM — mem0 memory extraction + digest pipeline
LLM_MODEL=minimax.minimax-m2.5
DIGEST_LLM_MODEL=minimax.minimax-m2.5
LLM_TEMPERATURE=0.1
LLM_MAX_TOKENS=2000

SERVICE_HOST=0.0.0.0
SERVICE_PORT=8230

# 关闭 mem0 匿名遥测（强烈建议，避免线程泄漏）
MEM0_TELEMETRY=false

LLM Selection

mem0 uses an LLM for two purposes:

• Memory extraction / dedup (LLM_MODEL) — called on every memory/add request
• Digest summarization (DIGEST_LLM_MODEL) — called by auto_digest.py to summarize diary entries

Both default to minimax.minimax-m2.5. Here's a comparison of supported models on AWS Bedrock us-east-1:

Model	LLM_MODEL value	Avg latency	Input	Output	Notes
MiniMax M2.5 ✅	minimax.minimax-m2.5	~2-3s	$0.30/1M	$1.20/1M	Default. Best cost (~3× cheaper than Haiku)
Claude Haiku 4.5	us.anthropic.claude-haiku-4-5-20251001-v1:0	~1s	$1.00/1M	$5.00/1M	Fastest; no patch required
DeepSeek V3.2	deepseek.v3.2	~2-4s	$0.62/1M	$1.85/1M	No clear advantage over MiniMax

WARNING

MiniMax requires a patch. Before using any minimax.* model, run:

python3 patch_minimax_support.py

See Known Issues for details.

Data Isolation

Two-dimensional isolation using user_id + agent_id:

• user_id — different users' memories are completely isolated
• agent_id — different agents of the same user manage memories independently
• Omitting agent_id allows cross-agent retrieval

Token Tracking

The service automatically tracks LLM token consumption for every /memory/add call via TrackedAWSBedrockLLM — a thin wrapper around AWSBedrockLLM that monkey-patches client.converse() to capture inputTokens / outputTokens / totalTokens from Bedrock Converse API responses.

How it works:

1. Before each memory.add() call, a global counter (protected by a threading lock) is reset.
2. Every Bedrock converse() call inside mem0 accumulates tokens into the counter.
3. After memory.add() returns, the counter snapshot is:
   • Returned in the API response as token_usage
   • Written to the daily audit log (audit_logs/audit-YYYY-MM-DD.jsonl) as a type=token_usage entry

Analyzing token usage:

# Daily summary
cat audit_logs/audit-$(date +%Y-%m-%d).jsonl | python3 -c "
import sys, json
from collections import defaultdict
records = [json.loads(l) for l in sys.stdin if l.strip()]
tok = [r for r in records if r.get('type') == 'token_usage']
total = sum(r.get('total_tokens', 0) for r in tok)
print(f'{len(tok)} calls, {total:,} total tokens')
"

TIP

Token tracking adds negligible overhead (a single dict write per Bedrock API call). It requires no extra configuration — it is always enabled when the service starts.