Letta

Executive Summary

Letta is a platform for building stateful AI agents with advanced memory systems. Originally the MemGPT research project from UC Berkeley's AI Research Lab (BAIR), it introduced the concept of virtual context management — using an OS-inspired memory hierarchy to give LLMs effectively unlimited context. The company emerged from stealth in September 2024 with $10M in seed funding led by Felicis, with participation from Founders Fund and YC.

---

How It Works

Letta's core innovation is a tiered memory architecture inspired by operating system memory management. Agents have:

• Core Memory — always in the LLM context window. Contains the agent's persona and key info about the current user. Editable by the agent in real-time.
• Archival Memory — a vector database for long-term storage. Agents can write observations and retrieve them later via search. Scales to unlimited size.
• Recall Memory — full conversation history stored in a database. Agents can search past messages rather than relying on a fixed context window.

The agent autonomously decides when to read/write across these tiers using tool calls. This creates a self-managing memory system where the LLM acts as its own memory controller — the key insight from the original MemGPT paper.

Recent Developments

Letta has expanded beyond pure memory into a full agent platform, and as of June 2026 Letta Code is clearly the company's center of gravity:

• Letta Code (Dec 2025) — a memory-first coding agent that runs in your terminal, scoring 42.5% on Terminal-Bench and ranked #1 model-agnostic open-source agent
• Letta Code app (Apr 2026) — a desktop app for "deeply personalized agents that learn over time and work locally on your machine"
• Channels (beta) — connect agents to Telegram, Slack, Discord, WhatsApp, and Signal directly from Letta Code; the standalone LettaBot project was archived in May 2026 in favor of Channels plus Remote Environments and letta cron
• Context Repositories — git-based memory for coding agents (Feb 2026)
• Context Constitution (Apr 2026) — published principles for how agents should manage context to learn from experience
• Conversations API — shared agent memory across concurrent user sessions
• ADE (Agent Development Environment) — web UI for building and managing agents

---

Pricing

Pricing was restructured around Letta Code in 2026 — the former $200/month Max tier no longer appears:

BYOK (Bring Your Own Keys) works on all tiers with OpenAI, Anthropic, and OpenRouter keys — and Letta Code can also draw on existing external coding plans. Letta's own docs note that casual coders typically run ~$100/month+ in usage and power users often exceed $200/month.

---

Strengths

• Deep research pedigree — born from a well-cited academic paper, not hype-driven
• Model-agnostic — works with OpenAI, Anthropic, Mistral, and local models via Ollama
• Open source core — Apache 2.0 license, self-hostable, 23K+ GitHub stars as of June 2026
• Full Python and TypeScript SDKs with a proper REST API
• Memory transparency — developers can inspect and control agent memory state, unlike black-box alternatives
• Active development — regular releases, expanding into coding agents and context repositories

Cautions

• Complexity overhead — the tiered memory system adds latency and token cost from internal tool calls
• Stability concerns — some users report instability, especially with local models
• Scope creep risk — expanding from memory layer into full agent framework and coding agent puts them in competition with many more players
• Community perception — some view the pivot from pure memory research to agent platform as chasing hype, though the founders dispute this
• Local model support — while technically model-agnostic, reliability with smaller local models can be poor

---

What Developers Say

The Letta Code launch thread on Hacker News (December 2025, 83 points) captured both the appeal and the skepticism around memory-first agents:

"Void is the greatest ad for Letta. I'm interested to see if it's as good at coding as it is at posting." — ascorbic, Hacker News

"In my experience, 'memory' is really not that helpful in most cases. For all of my projects, I keep the documentation files and feature specs up to date... Maintaining the memory is a considerable burden." — tigranbs, Hacker News

"Context poisoning is a real problem that these memory providers only make worse." — danieltanfh95, Hacker News

Co-founder Charles Packer told Tessl the resemblance to Claude Code is deliberate: "we're not trying to innovate at the UI layer, we're innovating at the 'memory' (context engineering) layer."

---

Competitive Positioning

---

Bottom Line

Letta occupies a unique niche: it's the most serious open-source attempt at solving LLM memory as a first-class concern. The MemGPT paper was genuinely influential, and the tiered memory architecture remains the most principled approach to agent state management available.

The 2026 trajectory shows consolidation rather than sprawl: the standalone LettaBot was archived in May 2026 and folded into Letta Code as Channels, the desktop app shipped in April, and pricing was rebuilt around Letta Code. The company is converging on one product — a memory-first coding agent — rather than maintaining parallel bets.

Recommended for: Developers building agents that need to remember and learn over long time horizons — Letta remains the most mature open-source option, and the free tier with BYOK makes it low-risk to evaluate. Also worth a look if you want a Claude Code-style coding agent that accumulates project knowledge across sessions.

Not recommended for: Teams that just need a lightweight memory API for an existing agent (consider Mem0), or workflows where up-to-date docs and specs already serve as the source of truth — several developers report that self-managed memory adds burden and context-poisoning risk without clear payoff.

Outlook: The bet is that memory becomes the durable differentiator in coding agents before frontier providers ship equivalent native features. The Terminal-Bench result and steady 2026 shipping cadence are credible signals, but the memory layer — Letta's core differentiator — still risks being absorbed by the model providers themselves.