Letta
An open-source framework for building AI agents that keep their state and memory under your own control — self-hostable, built on academic research (MemGPT) — instead of relying on a third-party's hosted black box.
🔗 Visit Letta
Description
Some teams want an AI agent's memory managed for them; others want full control over exactly how it works and where the data lives — the same trade-off as choosing a managed database service versus running Postgres yourself. Letta is built for the second group: it's an open-source framework, born out of academic research at UC Berkeley, for building 'stateful' AI agents that persist memory and can be run entirely on infrastructure you control.
Letta grew out of MemGPT, a UC Berkeley Sky Computing Lab research project on giving language models persistent, self-editing memory, and launched as a company in September 2024. It's Apache 2.0 open source with nearly 24,000 GitHub stars and offers Letta Code, an agent coding tool that ranked #1 on the model-agnostic Terminal-Bench benchmark. Pricing spans Free (3 stateful agents, bring-your-own API keys), Personal Pro ($20/month, 20 agents with a remote sandbox), Developer API ($20/month base plus $0.10 per active agent per month, plus pay-as-you-go LLM costs), Teams Pro ($20/seat/month for collaboration features), and custom Enterprise with SSO and volume pricing.
💬 Our review
The short version: Letta's roots in real academic research (MemGPT out of UC Berkeley) and its genuinely open, self-hostable architecture make it the pick for teams that specifically want to own their agent's memory infrastructure rather than depend on a hosted third-party service.
That academic-to-product path is also worth understanding as a trade-off: Letta is more of a developer framework to build with than a plug-and-play API like Mem0, so expect more setup work in exchange for more control — it's not a drop-in replacement if you just want the fastest path to 'my agent remembers things.' Letta Code's #1 ranking on Terminal-Bench is a genuine, independently-run benchmark result rather than a self-reported number, which is a stronger signal than most competitors can offer. Pricing is granular (a base fee plus per-agent cost), which scales reasonably for a handful of agents but is worth modeling carefully if you're planning to run agents at real scale. For teams prioritizing self-hosted control and open-source transparency over agent memory, Letta is a strong, research-backed choice; for teams that just want the fastest managed integration without running their own infrastructure, Mem0's hosted API is the simpler path.
💰 Pricing
📊 Global score
🤖 AI-enriched data
Free : 3 agents à état, clés API à apporter soi-même. Personal Pro : $20/mois (20 agents, sandbox distant). Developer API : $20/mois + $0.10/agent actif/mois + coûts LLM à l'usage. Teams Pro : $20/siège/mois. Enterprise : sur devis.
Pros
Racines académiques réelles (MemGPT, UC Berkeley) plutôt qu'un pur produit marketing
Open source Apache 2.0, self-hosting complet possible, 24k+ stars GitHub
Letta Code classé #1 sur le benchmark indépendant Terminal-Bench
Cons
Framework à construire dessus plutôt qu'une API plug-and-play comme Mem0
Plus de travail de mise en place pour plus de contrôle
Pricing par agent actif à modéliser soigneusement à l'échelle