// comparison

AgentRAM vs Letta

Last updated: June 2026

At a glance

	AgentRAM	Letta
Product category	Memory API only	Stateful agent framework
Pricing model	Pay-per-operation	Per-agent subscription
Free tier	100 operations, no card	3 managed agents, BYOK
Cheapest paid	$5 for 50,000 operations	$20 per month for 20 agents
Memory architecture	Key-value with TTL and text search	Tiered: core (in-context), archival, recall
Includes agent runtime	No, brings your own agent	Yes, full runtime
Self-hosted	No, hosted only	Yes, open source
Model-agnostic	Yes, LLM-side of your app	Yes, multi-LLM support
MCP integration	Official npm package	Community contributions
Multi-agent shared memory	First-class shared namespaces	Per-agent state by default
Origin	Solo built, 2026	MemGPT research, UC Berkeley, 2023
Backing	Bootstrapped, solo developer	$10M seed (Felicis, Founders Fund, YC)

The key conceptual difference

Letta is "agent and memory tightly integrated as one product". You write Letta agents, they run inside Letta's runtime, and their memory is managed by Letta's tiered architecture. The agent and the memory are inseparable.

AgentRAM is "memory you can plug into any agent". You write your agent however you want, in whatever framework you want, and call AgentRAM's HTTP endpoints when you need to store or retrieve. Memory is decoupled from the agent.

Neither approach is universally better. They serve different goals.

When Letta is the right call

Pick Letta if any of these match:

• You're building a stateful agent from scratch, not adding memory to an agent you already have. Letta gives you the whole stack, you don't have to assemble it.
• You want OS-inspired memory hierarchy. Letta models memory as three tiers: core memory (always in context), archival memory (large vector store), and recall memory (conversation history). This is a powerful pattern and Letta is the reference implementation.
• You want agents that edit their own memory. Letta agents can autonomously decide what to remember and what to forget, including evicting things from core memory to archival.
• You need a coding agent specifically. Letta Code is their flagship product and currently ranks #1 on Terminal-Bench among open-source model-agnostic agents.
• You need self-hosted deployment. Letta is open source and you can run it on your own infrastructure.
• You want the agent and memory tightly integrated as one product with a single mental model.

When AgentRAM is the right call

Pick AgentRAM if any of these match:

• You already have an agent. Maybe it's LangChain, CrewAI, MCP-based, OpenAI Assistants, custom code. You don't want to rewrite it on top of a new framework; you just need persistent memory.
• You don't want to adopt a whole new agent runtime. AgentRAM is an HTTP API. Bring whatever agent you have.
• You want framework independence. If you switch agent frameworks tomorrow, your memory layer is unchanged. With Letta, switching frameworks means leaving Letta.
• You're MCP-first. AgentRAM ships an official MCP server (npm install agentram-mcp) with 10 tools that work in Claude Desktop, Cline, Cursor, and any other MCP client.
• You want pay-per-operation pricing instead of per-agent. If you're building one large agent or running thousands of small ones, per-operation pricing tracks actual usage.
• Multi-agent shared memory is core to your use case. AgentRAM's shared namespaces let multiple agents read and write the same memory pool with built-in auth.

How to decide

The honest test: are you building an agent, or do you already have one?

If you're starting from scratch and want a complete stateful-agent platform with sophisticated memory architecture baked in, Letta is purpose-built for that. It's mature, well-funded, and the team has serious research credibility from the MemGPT work.

If you've already built an agent and just need a place to put memory, AgentRAM is the simpler, cheaper, more framework-neutral choice. You keep your existing agent and add HTTP calls for storage.