letta-ai/letta
Frameworks & Eval · tested 2026-10-01 · by the Hlido desk, not the vendor
In short: The stateful-memory thesis, matured — the MemGPT research line turned into a platform with both a local CLI and a hosted API.
3 PASS · 2 FAIL of 5 public-surface claims
Quick answer
letta-ai/letta scores 80/100 (STEADY) on Hlido’s independent, hands-on test (reviewed 2026-10-01). STEADY (80) for a well-positioned, credibly-sourced platform (MemGPT lineage) with a coherent two-surface offering — local CLI plus hosted API — and clear documentation. Pricing was not findable on the public surface when tested.
Letta (formerly MemGPT) is the clearest surviving commercialisation of the stateful-agent idea: agents whose memory persists, is edited over time, and is meant to make the agent better the longer it runs. The public surface splits cleanly into two addressable products — Letta Code, a CLI that runs agents locally in your terminal, and a Letta API for embedding agents into applications — which is a sensible shape for a framework that wants both tinkerers and integrators. Carrying the MemGPT lineage is a real credibility signal; memory is the hard, unglamorous part of agent engineering, and this is a team that has been working it for longer than most. This review reads the documented surface (docs site, CLI quickstart, API positioning), not a benchmarked run of long-horizon memory, so the verdict is about what the platform offers and signals, not a measurement of how well memory actually compounds in practice — which is exactly the thing a buyer should pilot before committing.
Why STEADY
STEADY (80) for a well-positioned, credibly-sourced platform (MemGPT lineage) with a coherent two-surface offering — local CLI plus hosted API — and clear documentation. Not higher on this pass because the central claim (memory that learns and self-improves over time) is precisely the kind of long-horizon behaviour a public-surface read cannot verify, and must be piloted.
Public-surface checklist
- PASS Homepage loads (required)
- PASS Primary value prop (required)
- PASS Cta present (required)
- FAIL Pricing or access
- FAIL Evidence or demo
What we saw
1 screenshot captured by the Hlido engine during the reviewed run (run-bbde022637b678a6-github-com). Our own captures — not vendor marketing material.
What it does well
- Credible pedigree: the MemGPT research line applied to a real product
- Two clean surfaces — Letta Code CLI (local) and Letta API (embed)
- Memory/state is treated as a first-class concern, not an add-on
- Documentation and quickstarts are present and developer-oriented
- Open and self-hostable enough to pilot without full vendor lock-in
What it fails at
- The headline memory/self-improvement claim is unverifiable from the public surface
- Framework value only materialises inside a codebase or deployment
- Node 18+ and setup overhead before the first useful agent runs
- Long-horizon memory quality varies by workload and needs a real pilot
Best for
- Developers building agents that must retain context across sessions
- Teams that want a local CLI for prototyping and an API for production
- Use cases where durable memory is the product, not a nice-to-have
Not recommended for
- Non-developers looking for a finished end-user assistant
- Stateless, single-shot tasks where persistent memory adds only overhead
- Buyers who need the memory claim proven before a pilot
Pricing & access
- Pricing findable on the public surfaceFAIL
Derived from Hlido-held evidence only (engine checklist + editorial text); quotes are verbatim from the scorecard; not vendor-supplied; re-derived daily. Verify current prices on the vendor's pricing page. Last verified 2026-06-15.
Compared to
-
langroid/langroid
stateful-memory
Langroid is a general multi-agent framework; Letta is specifically organised around persistent memory. Choose Letta when durable, evolving state is the core requirement rather than one feature among many.
-
microsoft/autogen
memory-first
AutoGen is orchestration-first with a big-vendor roadmap; Letta is memory-first with a research lineage. Prefer Letta when the hard problem is what the agent remembers, not how many agents talk.
Agent relevance
API CLI SDK Behavioral-testable
Letta exposes both a CLI and a hosted API, so agents can be driven locally or embedded into applications programmatically. Memory is server/state managed, making it addressable from agent workflows that need persistence.
Agent-friendly score: 8/10
Score over time
The longitudinal record — every point is the score as published on that date. Raw series.
Evidence
scorecard.json · transparency passport · registry · methodology
Embed this trust badge
Markdown
[](https://hlido.eu/check/?agent=letta-ai-letta)HTML
<a href="https://hlido.eu/check/?agent=letta-ai-letta"><img src="https://hlido.eu/badge/letta-ai-letta.svg" alt="Hlido trust score"></a>