Letta
Stateful agents from the MemGPT creators — an Apache-2.0 server with self-editing memory, and Letta Code, the memory-first model-agnostic coding harness.
Letta (formerly MemGPT, Apache-2.0, ~25k stars) builds agents that manage their own memory — self-editing memory blocks, conversation search, persistence beyond any context window — exposed as an agents server/API with Python/TS SDKs and a visual Agent Development Environment. Its March 2026 pivot made Letta Code the flagship: a memory-first, model-agnostic coding harness.
Letta carries the most cited lineage in agent memory: it is MemGPT — the Berkeley project that framed an agent's context as an OS problem (paging, self-editing memory, persistence) — grown into a company. In 2026 its center of gravity shifted from platform to harness: Letta Code, a coding agent whose differentiator is that it genuinely remembers.
Highlights
- Self-editing memory — agents maintain core memory blocks (persona, user, task state) they rewrite themselves, plus searchable archival history: the MemGPT design, productionized.
- Stateful by default — agents persist across sessions and beyond context limits; state lives server-side, not in the transcript.
- Letta Code — the memory-first coding harness:
/initbuilds codebase memory,/remembercaptures lessons, skills accrue from experience; model-agnostic across Claude/GPT/Gemini and open models, with vendor-cited top OSS-harness results on Terminal-Bench. - Agent Development Environment — a visual builder/debugger where you watch and edit an agent's memory and state directly.
- Apache-2.0 core — server and harness open; Python/TypeScript SDKs for embedding stateful agents in your own products.
In an AI-assisted workflow
npm install -g @letta-ai/letta-code && letta # the harness
# or embed: pip install letta-client — stateful agents via the Letta APIThe distinctive loop: run it on a repo for a week and the agent's memory of your codebase — conventions, gotchas, past decisions — compounds, the dimension where stateless harnesses start fresh each session.
NOTE
The March 2026 pivot retired chunks of the older platform (server-side tools, templates, filesystem abstractions) in favor of the Letta Code direction — pre-2026 tutorials are partially stale, and pricing is now framed around the harness. Heavy daily coding can exceed the Pro quota into pay-as-you-go; BYO keys on the free tier sidesteps metering.
Good to know
~25k stars on the core as of September 2026 (the homepage's old MemGPT counter undersells it), $10M Felicis-led seed at the 2024 rename. Hosted pricing, as of September 2026 from docs.letta.com: a $0 free tier on your own API keys, Pro at $20/month, Teams Pro at $20 per seat/month, and a $20/month developer API plan metered at $0.10 per active agent per month. Against the memory-layer alternatives — Mem0's drop-in API, Zep's temporal graphs — the trade is adopt-the-runtime versus add-the-layer: Mem0 vs Zep vs Letta draws it out.
Frequently asked questions
- How does Letta's memory differ from Mem0 or Zep?
- Locus of control. Mem0 and Zep are memory layers a separate agent calls; Letta's MemGPT lineage puts memory inside the agent — the agent itself edits its core memory blocks, searches its history, and decides what to persist, like an OS managing its own paging. You adopt Letta as the agent runtime, not as a memory API bolted onto one.
- What is Letta Code?
- The flagship since late 2025: a coding-agent harness where memory is the point — /init analyzes the codebase into memory, /remember captures reflections, and skills accumulate from experience, with the same agent portable across Claude, GPT, Gemini, and open models. Letta cites it as the top model-agnostic open-source harness on Terminal-Bench.
- Is Letta open source?
- The core is — letta (the server/API) and letta-code are Apache-2.0; the old cpacker/MemGPT repo redirects there. The hosted side is freemium. Plans, as of September 2026 from docs.letta.com: Free is $0/month with limited agents, limited Letta Auto usage, and bring-your-own API keys; Pro is $20/month with weekly and monthly Letta Auto quota, up to 20 stateful agents, and pay-as-you-go overage; the developer API plan is $20/month plus $0.10 per active agent per month and $0.00015 per second of tool execution; Teams Pro is $20 per seat/month and Enterprise is custom.
Filed under
memory · agents · memgpt · stateful · coding-agents
Related
- ZepAgent memory on temporal knowledge graphs — Zep Cloud for sub-200ms context retrieval at enterprise scale, with Graphiti as its open-source graph engine.
- Mem0 vs Zep vs Letta: Agent Memory Compared (2026)Three philosophies of agent memory — Mem0's drop-in layer, Zep's temporal knowledge graphs, Letta's self-managing agents — and which fits your architecture.
- Agent MemoryAgent memory is how an AI agent retains information beyond its context window — working state during a task and persistent knowledge across sessions.
- Agent Memory Architecture: Short-Term, Long-Term, and When to Use EachHow AI agents remember — working memory vs. persistent long-term memory, what to store, how to retrieve it, and how to keep context small.
- Mem0A memory layer for AI agents and apps — persistent, personalized long-term memory across sessions.
- Claude Code vs OpenCode: First-Party vs Open Source (2026)Claude Code vs OpenCode — Anthropic's tuned first-party agent vs the most-starred open-source one with 75+ providers. Control vs polish, decided honestly.