Overview
Letta (formerly MemGPT, 24k+ stars, Apache-2.0) is a stateful agents platform focused on agent memory, learning, and self-improvement, offering CLI, App Server, desktop, and web access. Graphiti (getzep, 30k+ stars, Apache-2.0) is a temporal context graph engine that builds and queries knowledge graphs where facts evolve over time, positioned as a dynamic memory alternative to RAG. We compare memory model, deployment form, query capabilities, and typical scenarios.
Unknown · Apache-2.0
Letta (formerly MemGPT) is an open-source framework for building stateful AI agents with advanced reasoning and transparent long-term memory. It allows you to visually test, debug, and observe agents.
Python · Apache-2.0
Graphiti is a temporal knowledge-graph engine for agent memory, helping systems continuously accumulate long-term context.
| Best for | Letta | Graphiti |
|---|---|---|
| Memory model | Agent stateful memory (advanced memory): README tagline is "can learn and self-improve over time." Inherits the three-tier architecture (core / archival / recall) from the MemGPT era. | Temporal context graph: nodes for entities, edges with validity windows for facts, episodes for provenance. Facts expire over time. |
| Deployment | CLI, App Server (self-hosted), desktop app (macOS / Windows / Linux), web (chat.letta.com), Letta Cloud, SDKs (TS / Python). Multi-device agent continuity. | Python library plus self-hosted. Bring your own graph database: Neo4j 5.26+, FalkorDB 1.1.2+, Amazon Neptune, Kuzu (deprecated). Requires LLM API key (OpenAI / Anthropic / Gemini / Groq). |
| Query and retrieval | Agents read and write memory via built-in tools (core memory editing, archival storage, recall retrieval). Designed for agent runtime calls. | Hybrid retrieval: semantic + keyword + graph traversal. Incremental updates without full recomputation. Sub-second latency. Bi-temporal tracking (when a fact was true plus when it was recorded). |
| Agent integration | Complete agent platform: CLI plus Server plus UI integrated. Agents can be hosted on desktop, server, or cloud. SDKs integrate into TS / Python apps. Slack, Telegram, Discord channel support. | Memory layer component: must be embedded into an agent framework (LangGraph, CrewAI, AutoGen). Shares DNA with the commercial Zep product (Graphiti is the open-source core of Zep). |
| Best fit | Stateful agent platform, self-improving learning, multi-device (CLI, desktop, web) access, users wanting a ready-made agent runtime. | Dynamic memory alternative to RAG, scenarios with evolving facts (CRM, customer service, personal assistants), existing agent frameworks needing better memory. |
| Metric | Letta | Graphiti |
|---|---|---|
| Stars | 24.6k | 30.5k |
| Forks | 2.6k | 3.1k |
| Language | Unknown | Python |
| License | Apache-2.0 | Apache-2.0 |
| Last commit | August 23, 2026 | September 1, 2026 |
Choose based on your primary workflow, language ecosystem, and integration needs. Review each project's documentation and recent GitHub activity before adopting it in production.
Pick Letta for a complete agent platform (CLI plus Server plus UI integrated). Pick Graphiti for upgrading the dynamic memory layer inside an existing agent framework. Both are Apache-2.0.
Letta is the new name for MemGPT. The MemGPT project was renamed Letta; active development has moved to letta-ai/letta-code (agent harness, interactive CLI, runtime), while letta-ai/letta is primarily a landing page plus documentation plus archived portions.
GraphRAG is batch-oriented with static summarization. Graphiti is incremental, with bi-temporal tracking and fact expiration. Graphiti handles evolving-fact scenarios better (CRM, customer service, personal assistants).
No. It supports Neo4j 5.26+, FalkorDB 1.1.2+, and Amazon Neptune. Kuzu 0.11.2 is deprecated; new projects should use Neo4j or FalkorDB.