====== Letta ======

**Letta** (formerly MemGPT) is an open-source platform for building stateful AI agents with persistent, transparent long-term memory. With **21.7K GitHub stars**, Letta pioneered the memory-first approach to agent design, where agents maintain identity, learn from interactions, and improve across sessions rather than resetting to a blank state.

{{tag>framework python agents memory stateful memgpt}}

===== Overview =====

Letta originated as MemGPT, a research project that introduced virtual context windows via memory tiers to overcome LLM context length limitations. In September 2024, MemGPT rebranded to Letta and expanded from a memory-augmented LLM prototype into a full platform for stateful agents. The core insight is that agents should have "lived experience" — storing interactions, building knowledge, and improving over time. Letta's architecture treats memory as the primary architectural concern, with agent reasoning built around persistent state rather than stateless prompt engineering.

===== Key Features =====

  * **Persistent Memory** — Core memory (identity), archival memory (long-term facts), and recall mechanisms
  * **Stateful Agents** — Agents persist across sessions, learning without stateless resets
  * **Memory Transparency** — Inspectable, debuggable memory with version control (git-backed)
  * **Model Agnostic** — Works with OpenAI, Anthropic Claude, local models via dynamic provider listings
  * **Letta Code** — Memory-first coding agent, #1 on TerminalBench (model-agnostic open-source)
  * **Agent Development Environment (ADE)** — No-code UI for building and testing agents
  * **REST API & SDKs** — Python and TypeScript SDKs, Vercel AI SDK provider
  * **Tool Integrations** — Composio, LangChain, CrewAI tool support
  * **Subagents & Skills** — Hierarchical agent composition with shared skill repositories

===== Architecture =====

Letta's memory-first architecture structures agents around a persistent state layer:

<mermaid>
graph TD
    A[LLM Reasoning Core: Observe - Reason - Act - Store] --> B[Core Memory: Identity / Goals / Context]
    A --> C[Archival Memory: Vector-stored Episodic and Semantic Facts]
    A --> D[Recall / Insertion Functions: Fetch and Prioritize]
    B --> E[Tools and Integrations]
    C --> E
    D --> E
    E --> F[Terminal]
    E --> G[Git]
    E --> H[Composio / LangChain]
</mermaid>

===== Code Example =====

Creating a stateful agent with persistent memory using Letta:

<code python>
from letta import create_client

# Connect to Letta server
client = create_client()

# Create an agent with persistent memory
agent = client.create_agent(
    name="research_assistant",
    system=(
        "You are a research assistant with persistent memory. "
        "Remember all interactions and build knowledge over time. "
        "Use your archival memory to store important facts."
    ),
    memory_human="User is a machine learning researcher.",
    memory_persona="I am a helpful research assistant that remembers everything.",
)

# Chat with the agent — it remembers across sessions
response = client.send_message(
    agent_id=agent.id,
    message="I'm working on a paper about transformer architectures.",
)
print(response.messages)

# Later session — agent remembers the context
response = client.send_message(
    agent_id=agent.id,
    message="What was I working on?",
)
# Agent recalls: "You mentioned working on a paper about transformer architectures"
print(response.messages)
</code>

===== Memory System =====

^ Component ^ Description ^ Use Case ^
| **Core Memory** | Fixed-size, always-in-context self-knowledge | Identity, goals, behavioral anchoring |
| **Archival Memory** | Vector-stored episodic/semantic facts | Long-term recall, knowledge accumulation |
| **Recall Functions** | Fetch/prioritize memories before LLM calls | Context optimization, relevance scoring |

Letta Filesystem achieves **74.0%** on the LoCoMo benchmark via simple file-based histories, demonstrating that effective long-term memory does not require complex architectures.

===== Evolution from MemGPT =====

  * **2023** — MemGPT paper introduces virtual context windows via memory tiers
  * **Sep 2024** — Rebranded to Letta, expanded to full agent platform
  * **Oct 2024** — v0.4-0.5: Composio/LangChain/CrewAI tools, dynamic LLM providers
  * **Dec 2025** — Letta Code: stateful coding agent, #1 on TerminalBench
  * **Mar 2026** — Next phase: git-backed memory, subagents, cross-provider deployment

===== References =====

  * [[https://github.com/letta-ai/letta|GitHub Repository]]
  * [[https://www.letta.com|Official Website]]
  * [[https://docs.letta.com|Documentation]]
  * [[https://github.com/letta-ai/letta-code|Letta Code]]

===== See Also =====

  * [[langchain]] — General LLM framework (tool integration)
  * [[agno]] — High-performance agent runtime
  * [[pydantic_ai]] — Type-safe agent framework