Research teams, advanced AI engineers building genuinely long-running agents, anyone implementing the MemGPT pattern in production.
Teams that need a quick agent SDK (use LangChain or CrewAI); applications that don't need persistent agent state.
What is Letta?
Letta (formerly the MemGPT project from UC Berkeley) is an open-source framework for building stateful AI agents with persistent memory and the ability to self-edit their own context window. Raised $10M seed in 2024 from Felicis. Distinct from Mem0: Letta is a full agent framework with memory baked in, not just a memory layer for other frameworks.
Key features
Integrations
What people actually pay
No price data yet — be the first to share
No price data yet for Letta. Help the community — share what you pay (anonymized).
The MemGPT pattern as a real product
Letta (formerly MemGPT) implements the self-editing-context pattern for stateful AI agents in a usable framework. More research-flavored than Mem0; the right pick for teams that want full agent state, not just memory.
Letta's lineage matters: it's the production version of the MemGPT research from UC Berkeley, which introduced the idea of LLMs that manage their own context window — promoting important facts into core memory, archiving less-relevant context to long-term storage, and self-editing what stays in active context across turns. As a framework, Letta is genuinely different from "agent SDK + memory layer" approaches like LangChain + Mem0; the agent itself is stateful in a deeper way.
The Agent Development Environment (ADE) is the underrated feature — visual debugging of agent state, memory, and reasoning across turns. For research teams and advanced AI engineers building genuinely long-running agents, this is a real productivity win. The trade-off is learning curve: Letta requires understanding the MemGPT pattern, which is more conceptual overhead than "give my agent a memory module."
Evaluate Letta if you're building genuinely long-running stateful agents (research assistants, persistent companions, autonomous workers) and want full agent state managed natively. Use Mem0 if you just need a memory layer for an existing agent framework. Skip if your agent is short-lived or stateless — Letta is overkill.
Research teams and advanced AI engineers building long-running stateful agents — research assistants, autonomous workers.
Teams that want a simpler agent SDK (use LangChain or CrewAI) or applications without genuine state requirements.
Written by StackMatch Editorial. StackMatch editorial reviews are independent analyst commentary, not user reviews. We have no affiliate relationship with this tool. See user reviews below for community perspective.
Before you buy Letta
Vendors don't tell you about their competitors. We do — with verdicts attached when we have them.
What Letta actually costs
Sticker price isn't the real cost. We add implementation, training, and a probability-weighted lock-in penalty.
When to negotiate Letta
Vendor sales pressure is non-uniform — quarter-close, year-end, and post-funding-round are your high-leverage windows.
Strong negotiation window. Reps will push for end-of-quarter signature. Don't move first — let them initiate the discount. Target 15-30% off list plus negotiated terms.
Take this to your sales call
11 questions vendor sales teams steer around — generated from Letta's pricing tier, lock-in profile, and editorial verdict.
- 1PRICINGLetta is starter-tier on the public site. What's the discount path for small-sized teams committing annually vs. monthly?
- 2PRICINGWhat overages or seat-overflow charges should we plan for? Show me the worst-case bill if our usage grows 2x in year 1.
- 3CONTRACTAuto-renewal: how many days notice is required to terminate, and what happens if we miss the window? Will you commit to a renewal-reminder email at 90 and 60 days?
- 4MIGRATIONData export: what's the complete spec — format, frequency, and what data does the export NOT include? After contract end, how long do we have read-only access?
- 5MIGRATIONImplementation runs 1-2 weeks. Who from your team is included by default, and who do we add at additional cost? Is a CSM assigned?
- 6FITIndependent analysis (StackMatch Editorial) flags this verdict: "The MemGPT pattern as a real product." How do you address this concern specifically for our use case?
- 7FITLetta is best for: Research teams and advanced AI engineers building long-running stateful agents — research assistants, autonomous workers.. We're [describe your situation]. Walk me through the failure modes if our profile doesn't match.
- 8FITConnect us with 2-3 reference customers at our company size in AI/ML — not the case-study list, customers who've been live for 18+ months and have churned at least one tool from your stack.
- 9INTEGRATIONLetta lists 4 integrations including OpenAI, Anthropic, Ollama. Which of OUR existing tools — bring our list — have you confirmed shipping integration with versus "on roadmap"? Show me the actual status.
- 10VENDORTrack record over the last 18 months: any pricing model changes, executive departures, layoffs, M&A activity, or material customer churn we should know about?
- 11VENDORIf you're acquired or shut down, what's the contractual continuity — source-code escrow, data portability, transition period? Show me the actual clause.
What to actually test in the demo
Vendor sales teams script demos to maximize close rate. Here's what they'd rather you not test — derived from Letta's lock-in profile and editorial verdict.
- 1PERFORMANCEBring YOUR data, not their demo data. Insist on running the demo workflow against a sample of your real records, files, or queries. If they refuse — that's a signal.
- 2PERFORMANCEEditorial flags: "The MemGPT pattern as a real product." Construct a demo scenario that directly tests this concern. Ask the rep to walk you through it in real time, not promise a follow-up.
- 3PERFORMANCELetta demo will be built around the happy path. Ask: "Show me what happens when [the most common failure mode in our context]" — make them improvise.
- 4EDGE CASESPush the limits live: largest dataset, longest workflow, most users concurrent. Vendors prep demos for medium loads — your real-world usage might 10x what they show.
- 5EDGE CASESMobile and offline behavior: how does Letta degrade on slow connections, on iPad, in airplane mode? Test in the demo if your team uses these surfaces.
- 6PRICINGFind the upgrade triggers. Which features force a paid plan? Which usage limits trigger overage? Get the rep to demo your team hitting each cap.
- 7INTEGRATIONVendors love their integration logo wall. Test the actual depth: pick the 2-3 (OpenAI, Anthropic-style) integrations you depend on most, and ask the rep to demo a real two-way data sync, not a marketing screenshot.
- 8INTEGRATIONAPI and webhook reality check: rate limits, payload size limits, retry behavior, auth refresh handling. Ask for actual API docs in the demo, not "we'll send those."
- 9MIGRATIONDemo the full data export workflow. Even with low lock-in, you want to see how clean the exit looks before signing.
- 10SUPPORTSubmit a real support ticket DURING the demo. Use the actual support channel customers use, not the rep's email. Time the response. This is your most honest data point about post-sale reality.
- 11SUPPORTAsk to be connected with a customer in the demo who you can email TODAY (not "we'll arrange a reference call next week"). The vendor's confidence in their references is a tell.
User Reviews
Be the first to review this tool