Letta

Tier 2: Persistent (agent-scoped)

Agents that manage their own memory (formerly MemGPT)

letta.com ↗ · ~38,000 (including MemGPT) GitHub stars · $10M (emerged from stealth)

Our take

Letta is the most ambitious project in this space. The idea that agents should manage their own memory (deciding what to remember, forget, and retrieve) is compelling on paper. And with GPT-4, it actually works. The problem is everything else. Local LLM support is described as "pure torture" in GitHub issues. Every memory self-edit breaks prompt caching, adding latency and cost. And you can't just use the memory component. You adopt the whole agent platform or nothing. If you're building a research prototype with GPT-4 and want to push boundaries, Letta is worth trying. If you need memory in production by next quarter, start elsewhere.

How it works

Treats the LLM context window like an OS treats RAM. Agents actively manage their own memory, moving data between in-context "core memory" (working set) and external "archival memory" (long-term storage). The agent decides what to remember and what to forget.

When to use Letta

  • + Building agents from scratch where memory is a first-class concern, not an afterthought
  • + Research and experimentation with agent memory architectures
  • + Teams using GPT-4 or Claude that want agents to self-manage what they remember

When to skip it

  • Teams that just need a memory layer for existing agents. Letta requires adopting its entire runtime
  • Anyone running local or open-source LLMs (90% error rates reported with non-OpenAI models)
  • Quick production deployments (steeper learning curve than alternatives)

What it does better than everything else

Agent-managed memory. No other tool gives the LLM itself the ability to decide what to remember, what to forget, and what to retrieve on demand. If you believe agents should own their memory lifecycle, Letta is the only serious option.

MCP support

Supported. Letta acts as an MCP host. Agents can consume external MCP servers. A community-built Letta MCP server also exists for managing agents externally. Not a pure MCP memory server.

Why MCP matters: 5 MCP servers every engineering team should run →

Pricing

Free

5,000 monthly credits, 2 agent templates, 1 GB storage

Paid

$20/mo (Pro): 20,000 credits, unlimited agents. Enterprise: custom. Self-hosted: free (Apache 2.0).

The gotcha nobody mentions

Local and open-source LLM support is unreliable. The memory self-management paradigm requires the LLM to make correct tool calls to manage its own memory. GPT-4 handles this. Most open-source models don't. One user reported 90% of requests resulting in errors with non-OpenAI models.

"Unless you're using MemGPT with OpenAI/GPT, getting MemGPT to work-as-advertised is pure torture."

Frequently asked

Is Letta the same as MemGPT? +
MemGPT is the research paper and original project. Letta is the company and production framework built from it. The repository has evolved significantly from the original MemGPT prototype.
Can I use just the memory component without the full agent platform? +
No. Letta's memory is tightly integrated with its agent runtime. You can't extract the memory system and plug it into a different agent framework.
Does Letta work with Claude or other non-OpenAI models? +
It supports Claude and other models with structured output / tool-calling capabilities. Results are best with GPT-4 and Claude. Smaller or local models often fail at the memory self-management tasks.

Related reading

Also in this space

Letta is Tier 2 — persistent (agent-scoped) memory.

It solves real problems for individual agents and users. But if your team needs shared, always-current memory that works across Cursor, Claude, and every other AI tool simultaneously — that's a different architecture entirely.

We're building that with Knowledge Plane. Join the beta →