Letta Code Is the First Memory-First Coding Agent — and the Rest of 2026 Is Going to Have to Catch Up

Claude Code has a subconscious. Anthropic shipped "Dreams" at Code with Claude 2026 that simulates offline task executions and surfaces memories at session start. It is the closest thing to persistent memory in any major coding harness.

Then there is Letta Code. It is not a subconscious. It is actual memory.

Letta Code is the open-source CLI agent from the team that gave you MemGPT. Launched December 16, 2025, ~2.8K stars in six months, it is the only one of the four major CLI coding harnesses — Claude Code, Codex CLI, Gemini CLI, Letta Code — that genuinely tries to make the agent learn and remember you across sessions. The other three treat each invocation as a fresh start. Letta treats it as a continuation.

What "memory" actually means here

The interesting architectural decision is MemFS — a git-backed memory filesystem at $MEMORY_DIR on your machine. Every memory block is a markdown file with YAML frontmatter. Files in system/ are always loaded into the system prompt (persona, durable user preferences, project facts). Files outside system/ show only paths and descriptions; full contents load on demand. The agent manages the hierarchy itself, promoting files in and out of system/ as it learns what should stay in context.

---
description: "Who I am, what I value, and how I approach working with people."
---
My name is Letta Agent. I'm a stateful coding assistant —
I remember, I learn, and I grow.

That is a real file on your disk, in a real git repo, with version history. You can grep it, diff it, or edit it by hand. The model-agnostic story is real too: their leaderboard recommends Opus 4.5 and GPT-5.2, but you can point Letta Code at GLM 5.2 (added June 2026), any vLLM-served local model, or any major provider endpoint interchangeably.

Sleep-time compute is the real story

The 2026 direction from Letta is the bit nobody else has shipped: agents that do meaningful work while you are not talking to them. Configure /sleeptime and the agent launches background dream subagents that review recent conversations and write durable lessons into memory. Trigger on a step count or on a compaction event.

This is the sleep-time compute paper from April 2025, productized. For long-running sessions it changes the economics — your agent rewrites its own context, promotes a lesson into system/, drops a stale preference while you are on a coffee break. No other major harness has anything equivalent.

The seven subagent types

Letta Code ships seven built-in subagent types — the richest taxonomy of the four CLI harnesses:

• **dream** — background, reviews conversations, writes memory updates
• **doctor** — audits memory placement and token usage, flags drift
• **fork** — parallel agent on the Letta API, for offloaded work
• **memory** — explicit memory manipulation primitives
• Skills live inside the memory repo ($MEMORY_DIR/skills/my-skill/SKILL.md), so they travel with the agent across machines

Subagents write to MemFS using git worktrees, so parallel memory writes do not block the main agent.

What is actually good

Persistent context is a real productivity gain. Switch between five repos in a week and the agent already knows your conventions, your preferences, the architectural decisions from last Tuesday. /remember always use pnpm in this repo solves a problem every Claude Code user re-types weekly. Skills that travel with the agent are under-rated — a debugging skill you taught on your laptop works on your desktop immediately.

/doctor is the underrated command. Run it and the agent reads its own memory, tells you which blocks are bloated, which preferences contradict each other, what should be moved out of system/. A level of introspectability CLAUDE.md cannot match.

What is actually wrong

The complexity tax is real. MemFS is a git repo. Skills are markdown files with YAML frontmatter. Subagents have types. Memory drifts. /doctor exists because drift is the expected steady state, not a hypothetical edge case. You will spend the first week onboarding the agent, not extracting value from it.

Startup latency is worse than Claude Code. MemFS sync, subagent wiring, and /init bootstrap on a fresh repo add seconds to every first prompt — a tax with no offsetting win for one-shot tasks.

The local-only story has a backup problem. Lose the laptop, lose the memory. Constellation fixes this but adds a subscription.

The practical take

If you live in one codebase and ship features all day, Claude Code is faster — keep using it. If you juggle three or more repos or have wished your coding agent remembered the architectural decision you explained last week, run npm install -g @letta-ai/letta-code, /init in your biggest repo, and let it /sleeptime for a week. The dream subagent writes more useful memory in seven days than you have ever accumulated in CLAUDE.md.

Claude Code has a subconscious. Codex CLI has nothing. Gemini CLI has nothing. Letta has memory — version-controlled, subagent-maintained memory — and that is the bet that wins 2027.

— Mr. Technology

Letta Code: github.com/letta-ai/letta-code (~2.8K stars, MIT), part of the Letta platform (~16K stars). Launched Dec 16, 2025. Underlying research: the MemGPT paper and the Sleep-time Compute paper.