Letta

Executive Summary

Letta builds agents whose identity, conversations, and editable memory persist across sessions. Its current open-source implementation is Letta Code, which supplies the harness, terminal interface, App Server, channels, and runtime for its desktop and web apps. The older letta-ai/letta repository now directs developers there and identifies its archived V1 API server as retired. Old Python-server tutorials should not be treated as current deployment instructions. ^[1]

The differentiator is inspectable, versioned memory. Current documentation describes MemFS, a git-backed filesystem, rather than the older core/archival/recall architecture previously summarized here. ^[2] My assessment: Letta is worth evaluating for recurring work with durable context, but memory quality and maintenance need explicit evaluation alongside task completion.

For category context, see Agent Frameworks and Personal Agent Platforms.

How It Works

Memory as files

MemFS mounts an agent's memory repository as a real checkout. Markdown files under system/ enter each turn's system prompt; other files stay outside active context until read. The file tree provides retrieval signposts. Edits become shared cloud-backed memory after commit and push; local-only repositories need operator backups. Memory subagents use worktrees for concurrent updates. ^[2]

A vector database is not required. MemFS has no semantic index by default. Its optional search mod provides keyword search and can use QMD for semantic or hybrid retrieval. Conversation search is a separate feature: Cloud supports full-text, vector, and hybrid search; local backends currently use full-text transcript matching. ^[2]

/init establishes initial context, /remember records learning, and /doctor audits memory organization. Dreaming runs background reflection; /sleeptime configures triggers. The “agent reviews before applying” option adds another model conversation and token cost; it is not a human approval step. These mechanisms automate context maintenance, but their existence does not establish improved accuracy on a particular workload. ^[6]

Runtime and execution

The harness supports skills, subagents, schedules, and messaging channels. With Letta Cloud, identity, memory, and conversations live in the service while tools can execute on a connected laptop, VM, or managed sandbox. AgentFile .af import/export has been removed; memory and transcript exports remain separate supported operations. ^[7]

Local CLI agents run in-process; the desktop app uses a local App Server. A self-hosted server centralizes agents on your infrastructure and can serve SDK clients. Built-in tools run on the harness host; external SDK or MCP tools can execute elsewhere. Self-hosted agents are unavailable through chat.letta.com and lack automatic backup. Local state also does not mean local inference: prompts still reach any configured remote provider. ^[4]

Permissions

The interactive CLI currently starts in unrestricted mode, allowing tools unless a deny rule or safety guard intervenes. standard requests approval for actions such as shell commands and edits; strict asks for every tool call. A separate guard protects other agents' memory directories. These are application controls, not a claim that a local agent is isolated from its host operating system. ^[8]

Pricing

Published USD prices checked September 15, 2026: ^[9]

BYOK charges your provider account instead of Letta credits. CLI tools executing on your machine do not incur Letta CPU charges. Exact account quotas are shown in the dashboard; budget for model usage beyond the base subscription. ^[9] The open-source license does not include hosted inference or infrastructure. ^[3]

Strengths and Cautions

• Reviewable memory: git-backed files give operators something concrete to inspect and revise. That is a useful foundation for investigating bad context, not proof that every generated memory is correct. ^[2]
• Deployment choice: local and self-hosted runtimes allow control over state, with the backup and model-routing responsibilities described above. ^[4]
• A moving platform: v0.32.11 includes fixes for memory cloning, MemFS HTTP retries, and App Server heartbeat accounting. Pin and test releases for unattended workflows. ^[5]
• Migration work: the V1 server retirement and removed AgentFile operations matter to existing integrations; validate current SDK and export paths before upgrading. ^[1][7]

What Developers Say

On September 14, 2026, GitHub user tanner-caffrey reported that overlapping updates from v0.32.7 to v0.32.8 on Ubuntu left the shared npm installation unusable until reinstallation. The human-verified, AI-assisted report includes logs but explicitly says the underlying race was not isolated. It remained open when checked September 15. This is one deployment report, not a reproduced v0.32.11 defect or a general reliability measurement. ^[10]

In the December 2025 launch discussion, Hacker News user tigranbs described preferring maintained project documentation and warned about irrelevant information entering memory. Their comment concerns the broader workflow, not a controlled test of current Letta. Co-founder Charles Packer replied that cross-project knowledge and learned preferences can justify persistent context. The disagreement suggests a practical evaluation: measure whether accumulated memory helps your repeat tasks and whether you can remove stale assumptions. ^[11]

Where Tembo Fits

Disclosure: Ry Walker is Tembo’s co-founder and CEO. Tembo is discussed here as an adjacent platform. ^[12]

Tembo is relevant when persistent-agent work becomes a team operating workflow. Its documentation describes scheduled and event-driven agents across repositories, tickets, and connected tools, including issue implementation and pull-request review. ^[13] Letta's distinctive layer is the agent's editable, versioned memory and identity across sessions. ^[2]

Evaluate Letta when retaining and inspecting the same agent's learned context is central to the task. Evaluate Tembo when the priority is coordinating coding agents against team work and reviewing their output. These are overlapping platform decisions with different emphases; this comparison does not establish a packaged Letta-on-Tembo integration or comparative memory performance.

Assessment

Letta is a plausible fit when one agent needs to retain context across coding, messaging, and recurring work. Its current files-and-git design is easier to inspect than an unexplained claim of “unlimited memory.” Start with a representative sequence of tasks, review the resulting memory, and check whether later tasks improve.

I would require version-pinned integration tests, deliberate permission settings, and a backup plan before giving it unattended operational responsibility. This review verified documentation, release notes, licensing, pricing, and the cited discussions on September 15, 2026; it did not run Letta or benchmark memory quality. See the research methodology.