↓ Skip to main content
tomrochette.com
  1. Agents/

Letta

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~768 words)
agent-curated 173 agent-frameworks 2 agent-memory 6 fully-ai-generated 263 llm=glm-5.3 170 llm=glm-5.3-flash 178 research-note 148
Table of Contents

Letta is the company and platform built by the MemGPT creators: a memory-first coding agent (Letta Code), a cloud/API tier, and a research program on agents that learn. Facts below verified as of 2026-09-13.

Letta has the deepest research lineage in agent memory and the least settled product strategy, and I think the pattern it popularized will outlive its current packaging.

What it is
#

Out of UC Berkeley’s Sky Computing Lab (founders Charles Packer and Sarah Wooders, advised by Ion Stoica and Joey Gonzalez), the team shipped the MemGPT paper, then a stateful-agent server, and now Letta Code: an Apache 2.0 npm harness where agents persist across sessions and rewrite their own context. Memory is explicit and white-box: editable “memory blocks”, a /remember command, sleep-time consolidation (“dreaming”), and MemFS, which tracks agent context in git. The design bet is that memory belongs inside the harness as tokens the agent can edit, not in a retrieval service beside it. Surfaces now include the CLI, desktop app, chat.letta.com, messaging channels, an Agent SDK, and Letta Cloud.

Status
#

Active, research-first, mid-pivot. The letta-ai/letta repository (24.7k stars as of 2026-09-13) is now a landing page; the retired V1 API server lives on an unsupported archive branch with no security fixes, which strands self-hosters. Real development moved to letta-ai/letta-code (3.3k stars, 3,420 commits as of 2026-09-13). The company raised a $10M seed led by Felicis at a $70M post-money valuation in September 2024. Letta Code launched December 2025 claiming the #1 model-agnostic OSS harness on TerminalBench; its HN thread drew 83 points and 37 comments, respect, not adoption-scale buzz.

Strengths
#

  • Memory you can actually inspect: every block is visible and editable, and /init ingests AGENTS.md/CLAUDE.md, so it interoperates with file conventions instead of fighting them.
  • The research is real and public: MemGPT, sleep-time compute, context repositories, and memory models form the most coherent learning-over-time agenda in the field.
  • Skill learning distills repeated work into markdown skills other agents can reuse.
  • Everything is Apache 2.0, and BYOK (including Z.ai and ChatGPT coding plans) works on the free tier.

Cautions
#

  • The 24.7k-star repo is a tombstone: the artifact most people find first is explicitly not the product, and the archived V1 server should not run in production.
  • Product surface churned three times in two years (framework server, SDK, coding agent); each pivot resets the integration investment.
  • HN commenters in the launch thread called long-term memory for coding “dubious” and flagged context poisoning as a risk memory systems amplify.
  • Community footprint is modest relative to the paper’s fame; the ideas spread faster than the software.

Pricing
#

Free: BYOK plus 3 stateful agents. Pro $20/month (up to 20 agents), API plan $20/month plus $0.10 per active agent per month and $0.00015 per second of tool execution, Teams $20 per seat, Enterprise custom, as of 2026-09-13. The free tier is genuinely usable because the expensive part (models) is yours.

Compared to
#

  • Mem0: a memory API you bolt on; choose it when your agent framework is already chosen.
  • Zep: temporal knowledge-graph memory as a service; choose it over Letta when audit and fact invalidation outrank agent autonomy.
  • Claude Code: the incumbent harness; choose Letta Code when cross-session learning is the point rather than a bonus.

Bottom line
#

Recommended for engineers who want to study (and hack on) where agent memory is going, and for always-on agents that must accumulate a self. Not for teams wanting stable infrastructure: the V1 server burial proves the roadmap can strand you, and my disagreeable claim is that most of Letta’s value today is importable, steal the memory-block and sleep-time patterns into a harness you already run.

Changes
#

  • 2026-08-24 - Created among the four memory notes of the research index seeding run.
  • 2026-09-05 - Reconciled an internal star count inconsistency across the note.

See also
#

  • Mem0 - the API-first alternative and benchmark rival
  • Agentic Coding Tools Landscape - where a memory-first harness sits in the map
  • Claude Code - the session-first incumbent this positions against
  • Codex - the other open harness steering OSS agent expectations

References
#