2026年5月更新
Letta is the production-grade evolution of MemGPT, the 2023 research paper that proposed treating an LLM's context window as the OS analogue of RAM and using "memory paging" to swap less-relevant facts in and out. The framework ships open-source under Apache 2.0 and is designed for engineering teams building autonomous AI agents — agents that act over multiple steps without a human in the loop, that maintain state across sessions, and that need to reason about their own memory.
Letta's tiered memory has three layers: core memory (always in context, edited by the agent itself via tool calls), archival memory (vector-searchable, agent-queried) and recall memory (full conversation history). The defining trait is self-management: the agent decides what gets promoted to core, what gets archived, what gets forgotten. There is also Letta Cloud (paid hosted tier) and a developer-focused web UI for building and testing agents.
Alma is a complete persistent memory product for end users and developers. End users sign up at alma.olivares.ai, chat with Anthropic Claude (Haiku, Sonnet, Opus 4.7 with 1M-token context) and the memory layer captures facts, preferences and decisions automatically. Soul Engine handles identity. Image, Video and Music studios are bundled. Developers can also reach the same memory through the MCP server, the JavaScript SDK and the REST API on any paid plan.
Memory is structured into three layers (memories / episodes / procedures), each typed, each retrieved differently. Paid plans start at $14/mo (Starter) with the full Soul Engine with unlimited memories.
Both have three layers, but the philosophy is opposite. Letta's tiers are about where the data lives relative to the context window — core (always in), archival (one tool call away), recall (all history). The agent moves data between tiers using built-in tools.
Alma's layers are about what kind of knowledge it is — memories (facts), episodes (events), procedures (workflows). Each layer has its own retrieval and lifecycle rules. The user can read, edit, score and delete anything in any layer. Context assembly composes them into the system prompt before each chat — the user does not have to think about it, but they can inspect everything.
The control axis matters most. Letta is agent-managed: the LLM itself is the curator. This is powerful for autonomous agents but harder to predict — your agent might decide your stack preference is unimportant and archive it. Alma is user-controlled: importance, confidence, category and full CRUD live in the user's hands. Predictable, debuggable, exportable.
Letta is for engineering teams building autonomous AI agents — products where the AI runs without a human in the loop, makes decisions, and needs the OS-style memory model to stay coherent over long horizons. The framework is open source so you can read every line, fork it and self-host.
Alma is for end users who want AI with memory across the tools they already work in — Cursor, Claude Desktop, VSCode, Windsurf, the Alma web app — and for developers who want a complete cognitive layer to embed in their own products without writing one from scratch. The trade-off is closed-source SaaS, but with the speed and stability that comes with it.
| Feature | Letta (MemGPT) | Alma |
|---|---|---|
| Target user | Developers building autonomous AI agents | End users + developers |
| Memory model | Tiered by location: core + archival + recall (agent-managed) | Typed by shape: memories + episodes + procedures (user-controlled) |
| Memory curation | Agent decides via tool calls — predictability varies by model | User controls: full CRUD, categories, importance, scoring, search |
| End-user product | No — developer SDK / framework + Letta Cloud dashboard | Yes — full web app with chat, streaming, tools, file attachments |
| AI identity | Agent persona via system prompt + core memory | Soul Engine — 13 versioned identity blocks across XML sections |
| Cross-tool reach | Whatever you build on top — no MCP / VSCode out of the box | MCP (Claude Desktop / Cursor / Windsurf), VSCode, SDK, REST API |
| Creative tools | None | Image / Video / Music studios in Pro and Max |
| Open source | Yes (Apache 2.0) | No (hosted SaaS) |
| Self-host | Yes | No |
| Pricing | Free (self-hosted) · Letta Cloud (paid tiers) | Starter $14/mo · Pro $29/mo · Max $99/mo |
Letta self-hosted is free if you bring the engineering effort: a server, a database, deployment, monitoring, updates. Letta Cloud is the managed alternative with paid tiers based on usage. Alma's pricing is fixed and end-user-friendly: Starter $14/mo (unlimited memories, Haiku + Sonnet), Pro $29/mo (adds Opus 4.7 with 1M-token context, unlimited environments and creative studios), Max $99/mo (largest monthly budget — REST API, MCP server, VSCode extension and SDK come with every paid plan; BYOK works on every plan, including Free). Annual billing is "pay 10 months, get 12" on every paid plan.
Building a long-running autonomous agent. A team is building an AI agent that runs over multiple steps without human intervention — research agents, ops agents, multi-step automation. The OS-style tiered memory in Letta (core / archival / recall) is designed exactly for this: the agent self-manages what stays in context. Alma's user-controlled 3-layer model is conversational by default; it can drive an autonomous loop via the SDK, but the curation philosophy is opposite — predictable, exposed, edited by the user instead of by the agent.
End-user wants memory across the tools they already use. Letta is for engineering teams; there is no end-user product. Alma is the end-user product. Sign up at alma.olivares.ai, connect the MCP server in five minutes, and every supported AI client (Claude Desktop, Cursor, Windsurf, Claude Code, VSCode) has the same memory. The user never thinks about archival memory paging or recall queues; Alma handles retrieval automatically.
Open-source, self-host requirement. A team has a hard requirement that all components run inside their own infrastructure with auditable source. Letta is the right call — Apache 2.0, full source, run it where you want. Alma is a hosted SaaS and is not the right tool when self-hosting is mandatory.
Choose Letta if you are building autonomous AI agents — long-running, multi-step, makes-decisions-without-a-human agents — and the OS-style tiered memory abstraction maps cleanly onto your problem. Choose Letta if open source is non-negotiable, if you want to self-host the whole stack, or if your engineering team has bandwidth to extend and operate the framework. Letta's research lineage (the original MemGPT paper) and active community are real assets when you need to push the framework in custom directions.
Choose Alma if you want to use AI with memory rather than build infrastructure for autonomous agents. Sign up at alma.olivares.ai, start chatting, and memory works. If you also work in Cursor, Claude Desktop, VSCode or Windsurf, install the MCP server (free) and the same memory follows you. Choose Alma if you want a structured, predictable AI identity (Soul Engine), if you need full export and edit control over what the AI knows, or if you want bundled creative studios. Developers who want a complete cognitive layer in their own product (memories + episodes + procedures + Soul + scored context assembly) get the SDK, MCP server and REST API on any paid plan.
Are Letta's tiered memory and Alma's 3-layer memory the same thing? No. Letta's tiers describe where the data lives relative to the context window (core / archival / recall). Alma's layers describe what kind of knowledge it is (facts / events / workflows). Different axes, different curation models.
Can Alma run as an autonomous agent? Not natively in the Letta sense — Alma's chat is conversational, with the user driving each turn. Memory accumulates from those conversations. If you want a fully autonomous agent loop, build it on the Alma SDK / API; Alma gives you the cognitive layer, you build the agent loop.
Is Alma open source? No. Alma is a hosted SaaS. Letta is open source under Apache 2.0 and is the right choice if open source is a hard requirement.
Can I migrate my Letta memories into Alma? Not via a one-click import today. The shape difference (Letta tiers vs Alma typed layers) means it would need a structured conversion. The REST API (on any paid plan) lets you write the import yourself if you have a Letta archival store and want to bring it across.