Letta (MemGPT)
Open-source agentic framework (formerly MemGPT) with a REST API server providing agents persistent, structured memory management — enabling agents to remember users, facts, and history across unlimited conversations.
Score Breakdown
⚙ Agent Friendliness
🔒 Security
Self-hosted deployment provides full data control; default local dev config has no auth — must be hardened before network exposure
⚡ Reliability
Best When
You need agents with rich, durable memory that persists across sessions and can self-manage what to remember — especially when data privacy requires self-hosting.
Avoid When
Your agents are stateless, session-scoped, or you need a fully managed cloud platform without operational burden.
Use Cases
- • Build personal assistant agents that accumulate and recall user preferences, past decisions, and long-term context across many sessions
- • Create customer-facing agents that maintain persistent relationship context (purchase history, preferences, prior issues) per user
- • Deploy research agents that continuously update a structured knowledge base as they discover new information
- • Implement agents with tiered memory (in-context, archival, recall) that intelligently manage what to keep active versus store
- • Run self-hosted conversational agents for sensitive data use cases where cloud memory storage is not acceptable
Not For
- • Single-turn stateless tasks — Letta's memory architecture adds unnecessary overhead for simple request-response agents
- • Teams needing a fully managed cloud platform with SLA guarantees — self-hosted requires infrastructure management
- • Workflows requiring real-time sub-200ms responses — memory retrieval and context management add latency
Interface
Authentication
Self-hosted server can run with no auth (local dev) or with API key; Letta Cloud uses API key authentication
Pricing
Open-source Apache 2.0 for self-hosted; Letta Cloud is freemium with paid tiers for production use
Agent Metadata
Known Gotchas
- ⚠ Memory edit tools (core_memory_append, archival_memory_insert) are called autonomously by the LLM — agents can accumulate stale or contradictory memories without explicit cleanup logic
- ⚠ Archival memory search is semantic (vector similarity) — exact-match retrieval is unreliable; agents should not rely on it for structured key-value lookups
- ⚠ Self-hosted server requires PostgreSQL with pgvector extension for production — SQLite default is not suitable for concurrent multi-agent workloads
- ⚠ Agent persona and human persona strings have character limits — truncation without warning can cause subtle context degradation
- ⚠ Streaming responses are supported but memory writes happen post-stream — tools that depend on updated memory immediately after a streamed response may read stale state
Alternatives
Full Evaluation Report
Detailed scoring breakdown, competitive positioning, security analysis, and improvement recommendations for Letta (MemGPT).
Scores are editorial opinions as of 2026-03-06.