AI Agent Memory Comparison

Real data. Live GitHub stars. Updated March 2026. Find the right memory layer for your AI agent — from free open-source to enterprise cloud.

★ Star on GitHub Mem0 on GitHub → Letta on GitHub →

Last updated: March 28, 2026 · Data from GitHub API, Brave Search, official pricing pages

GitHub Stars & Pricing

System	GitHub Stars	Free Tier	Paid Plans	Open Source
Mem0 ★ 24K	★ ~24K	Free 10K memories/mo	$19/mo Starter $249/mo Pro	Open Core
Letta (MemGPT)	★ ~17K	Free Self-host only	$19/mo Starter $249/mo Pro	✓ Apache 2.0
Zep	★ ~4.3K	Free Hobby tier	$49/mo $249/mo Pro	✓ Apache 2.0
agent-memory ★ MCP	★ 0 (just launched)	✓ 100% Free No limits, no API key	✓ Free Forever No cloud, no account	✓ MIT License
Mem0 OpenMemory NEW	★ (in m0 repo)	Free tier Requires OpenAI key	$19+ /mo Sends data to cloud	Open Core

Feature Matrix

Feature	Mem0	Letta	Zep	agent-memory
MCP Support	✗ No	✗ No	✗ No	✓ MCP v3.2
Local-first	Cloud only	✓ Self-host	Cloud + Self	✓ JSON/SQLite
TTL Expiration	✓	✓	✓	✓ Built-in
Encryption	?	✓	✓	✓ AES
Multiple Backends	✗	✗	✗	✓ JSON/SQLite/Redis
REST API	✓	✓	✓	✓
SSE Streaming	?	?	?	✓
No Cloud Dependency	✗	✓	✗	✓ 100% local
LangChain Required	✗	✓	✗	✗

Memory Footprint (Tokens/Conversation)

Mem0

1,764 tokens

agent-memory

~2,000 tokens

Letta

~8,000 tokens

Zep

600,000+ tokens ⚠️

Source: Mem0 paper (2026). Zep's 600K+ token footprint vs Mem0's 1,764 is a 340x difference. agent-memory uses on-demand retrieval similar to Mem0.

💡 Key Insight — NEW: Mem0 OpenMemory vs agent-memory

Mem0 just launched OpenMemory MCP (mem0ai/mem0-mcp on GitHub) — an MCP server for coding agents (Cursor, VS Code, Claude). This validates the market but OpenMemory requires your OpenAI API key and sends data to Mem0's cloud. agent-memory is the only truly local-first, API-key-free MCP memory — your data stays on your machine. No cloud. No API key. No account.

When to Choose What

Scenario	Best Choice	Why
MCP-native (Cursor, Claude Code, OpenClaw)	agent-memory ✓	Only MCP v3.2 native option
MCP-native + no API key + offline	agent-memory ✓ ONLY	OpenMemory requires OpenAI key + cloud
Privacy-first, no cloud, no API key	agent-memory ✓	100% local, MIT license
Need Redis backend integration	agent-memory ✓	Built-in Redis support
Enterprise, need managed cloud	Mem0	$24M funded, production-grade
Need Letta agent framework	Letta	Full agent runtime built-in
Already using LangChain/LangGraph	Mem0 or Letta	Native LangChain integration

Quick Start

Install agent-memory

# Install
pip install agent-memory

# Use with MCP (Cursor, Claude Code, OpenClaw)
python -m agent_memory.mcp_server

# Or use directly
from agent_memory import Memory
m = Memory(storage="json", path="./memory.json")
m.add("User prefers dark mode", ttl="30d")
results = m.search("dark mode preference")

Connect via MCP tools

# MCP tools available:
memory_search    — Search memories
memory_add       — Add new memory  
memory_get       — Get memory by ID
memory_list      — List recent memories
memory_clear     — Clear memories

# Works with:
#   Cursor IDE (MCP integration)
#   Claude Code (MCP)
#   OpenClaw (MCP)
#   Any MCP-compatible client

agent-memory for coding agents

# AI coding agent MCP (Claude Code, Cursor)
python agent_memory/mcp_coding_agent.py

# Coding-specific tools:
memory_project_add   — Learn project structure
memory_code_fact    — Store API shapes, schemas
memory_decision     — Record architectural decisions
memory_session_resume — Resume work across sessions

📊 Market Context — March 2026

AI agent memory is becoming standard infrastructure. Mem0 hit 24K GitHub stars in 12 months. NVIDIA, Google, and OpenAI are all building agent memory layers. The market validates the category — but every major player requires cloud dependency or LangChain. agent-memory is the privacy-first, MCP-native alternative.