1. Should you try Letta? Overview

What it is	ai-productivity-data / Dev (agent-memory sublane)
Country / HQ	San Francisco, USA (MemGPT)
Best for	Agents that must remember users/preferences/project state across sessions
Check before buying	Watch out: Stateless one-shot LLM calls (→ raw SDK)
Pricing signal	Free tier + paid

2. Features What it does

Feature	What it does
Memory tiers	Core (in-context blocks, editable by agent), Recall (searchable conversation history in DB), Archival (vector store, unlimited cold storage) — Agent decides read/write/evict via tool calls; survives restart
Agent loop	Tool-calling loop, self-editing memory, context windowing, sleep-time consolidation — Not ReAct-only; MemGPT-style virtual context
Letta Code	Memory-first coding agent, desktop app (mac/win/linux Apr 2026), Channels (Slack/Telegram/WhatsApp/Discord/Signal May 2026), Terminal-Bench SOTA — Uses Letta Auto (managed model routing)
Skill learning (Dec 2025)	Agents extract reusable skills from trajectories, store in memory, reuse later (+21.1% TB1, +36.8% TB2) — Online learning, no fine-tune
Conversations API	Parallel conversations share agent memory; MemFS memory filesystem — Git-backed memory (late 2025): version/diff/rollback
Tools / MCP	Python tool bindings, remote MCP server support, scheduled runs, subagents, secrets mgmt — Client-side/remote MCP no server-exec charge
Surfaces	REST API (v1 agents/messages/tools/blocks/archives), Python SDK letta-client, TS SDK @letta-ai/letta-client, CLI, ADE (web UI), desktop app — Self-host parity with cloud core
Deploy	Self-host (Docker, Postgres/SQLite/Redis/ClickHouse/Turbopuffer), Letta Cloud sandboxes, BYOK all tiers — Apache-2.0 core; platform closed cloud
Enterprise	SAML/OIDC SSO, RBAC, increased quotas, private model deploy, dedicated support — No native APM/eval (pair with Langfuse/Braintrust)
No	No eval/CI-gate suite (≠ Braintrust/LangSmith), no drag-drop builder, no non-LLM APM — Scope = agent runtime+memory

3. Pricing Pricing (2026-10, USD)

Plan	Price / mo	Key limits & who it fits
OSS Self-host	$0 (Apache 2.0)	pip install letta[server], BYOK, pay LLM vendor only, no agent cap
Cloud Free	$0	≤3 managed stateful agents, BYOK, limited Letta Auto, ADE access
Pro	$20/mo	≤20 stateful agents, Letta Auto weekly+monthly quota, remote sandboxes, pay-as-you-go overage, image-gen credits
Max Lite	$100/mo	≤50 stateful agents, higher Auto quota
Max	$200/mo	Highest personal quota, early-access features, not for commercial deploy
API / Developer	$20/mo base	Unlimited agents + 0.10/activeagent/mo∗∗+∗∗0.00015/sec server-side tool exec + pass-through LLM tokens; client/remote-MCP tools free
Teams Pro	$20/seat/mo	Shared org agents, access control, Pro features per seat
Enterprise	Custom	SAML/OIDC, RBAC, private models, volume pricing, SLA

4. Integration Access & integrations

Access type	Details
SDK	Python letta-client, TypeScript @letta-ai/letta-client
REST	v1 /agents /messages /tools /blocks /archives /runs /mcp etc.
CLI	letta server, letta run, letta create-agent
Self-host	Docker, Postgres/SQLite, Redis, ClickHouse (traces), Turbopuffer (tool search)
Channels	Slack, Telegram, Discord, WhatsApp, Signal (Letta Code)
MCP	Exposes MCP server; consumes remote MCP tools
Models	OpenAI, Anthropic, Mistral, Gemini, DeepSeek, Qwen, Ollama/LM Studio (BYOK)
No	No webhook-only SaaS for non-devs, no REST “score text” endpoint

5. Reviews & Sources External reviews

Source	Rating	What it says
AI Pedia / AI Tools Atlas / AITrendTool	7.5–8/10; “memory-as-infra done right”, agent self-edits core blocks, Letta Code Terminal-Bench SOTA, Apache-2.0 self-host. Criticized: 3-agent free cap tight, memory governance (retention/deletion/sensitive data) is dev responsibility, no built-in evals, Max tier confusing for commercial use
Community (GitHub 23K★, Reddit/Discord)	Loved: crosses model swaps without losing state, MCP + Channels, skill learning. Disliked: learning curve on memory-block design, Postgres ops for self-host, Letta Auto overage surprise, not turnkey for no-code PMs
Aggregate	Best stateful-memory runtime for agents that must remember users/projects across weeks; wrong for one-shot stateless calls or teams wanting LangSmith-style eval gates out of box

6. Best for / Not for Fit

✔ Suitable	✘ Not suitable
• Agents that must remember users/preferences/project state across sessions • Long-running assistants (customer success, tutoring, career coach) • Model-agnostic stacks swapping GPT↔Claude↔Qwen without losing state • Memory-first coding agents (Letta Code learns repo over time) • Self-host OSS with Postgres-backed agent state	• Stateless one-shot LLM calls (→ raw SDK) • Teams wanting built-in evals/CI gates (→ Braintrust/LangSmith) • No-code drag-drop builders (→ Lindy/n8n) • Simple RAG+vector-store needs (→ Mem0/Zep cheaper) • LangChain-locked teams wanting zero new runtime (→ LangGraph)

7. Competitors How it compares

Competitor	How it differs	Best for
Agent runtime	LangGraph	LangGraph = graph orchestration + deploy (LangSmith); Letta = memory-OS + self-editing agents, model-agnostic
Agent framework	OpenAI Agents SDK, Microsoft Agent Framework	Vendor/eco-locked; Letta Apache-2.0 + BYOK + tiered memory
Memory-layer add-on	Mem0, Zep	Mem0/Zep bolt onto existing agents; Letta is full agent runtime with memory native
OSS observability	Langfuse	Langfuse traces/evals; Letta runs agents, pairs with Langfuse for obs
Eval/obs platforms	Braintrust, LangSmith	They observe/eval; Letta executes stateful agents (complementary, not direct)
Coding agent	Cline, Aider, Claude Code	Those are coding agents; Letta Code is memory-first variant with Letta runtime underneath
In-list overlap	Exa, Cody, JB AI, Continue, Aider, Cline, Mintlify, Braintrust, LangSmith, Q Dev(sunset), Phind(shutdown)	Dev lane adjacent (agent-memory vs coding/eval/docs)

8. Summary Verdict

Who should pick it	Agents that must remember users/preferences/project state across sessions
Who should skip it	Stateless one-shot LLM calls (→ raw SDK)