Mem0 — Memory Layer for AI Applications

What it is

Mem0 (pronounced 'memo') is a memory infrastructure layer for AI applications. It automatically extracts, stores, and retrieves relevant facts and preferences from conversations, giving AI agents persistent context across sessions without manual prompt engineering.

The project suits developers building chatbots, AI assistants, or agent systems that need to remember user preferences, project context, or learned behaviors between sessions.

How it saves time or tokens

Without persistent memory, developers stuff prior conversation history into every prompt, burning tokens on repetitive context. Mem0 replaces this pattern with a search-based retrieval system: only relevant memories are fetched per query, reducing prompt size and improving response accuracy.

How to use

1. Install Mem0 with pip install mem0ai.
2. Initialize the Memory object and add facts using m.add('fact', user_id='user').
3. Search memories with m.search('query', user_id='user') to retrieve relevant context for your prompts.

Example

from mem0 import Memory

m = Memory()

# Store user preferences
m.add('I prefer Python over JavaScript', user_id='alice')
m.add('My project uses PostgreSQL and Redis', user_id='alice')

# Retrieve relevant memories
results = m.search('what database does alice use?', user_id='alice')
print(results)
# [{'memory': 'Uses PostgreSQL and Redis', 'score': 0.95}]

# Multi-level memory: user, agent, and session scopes
m.add('Dark mode preferred', user_id='alice')
m.add('Always respond in bullet points', agent_id='support-bot')

Related on TokRepo

• AI Memory Providers -- compare memory systems like Mem0, Zep, MemGPT, and Letta
• Mem0 Deep-Dive -- detailed breakdown of Mem0 features and integration patterns

Common pitfalls

• Do not store entire conversation transcripts as memories. Mem0 works best with extracted facts and preferences, not raw text dumps.
• Memory search is semantic, not keyword-based. Queries like 'database' will match 'PostgreSQL' even without exact word overlap.
• For production use, configure a persistent vector store backend. The default in-memory store loses data on restart.