Definition

What it is

An open-source **OS-style memory management framework** for LLM agents (formerly **MemGPT**), built on the analogy that the LLM's context window is RAM and external storage is "disk" — the agent's runtime swaps memory blocks in and out of the context window like an operating system pages virtual memory. Letta exposes core memory (always-resident persona + user state), recall memory (searchable conversation history), and archival memory (large semantic store), each backed by configurable storage tiers including PostgreSQL, vector stores, and S3-compatible object storage.

Why it exists

Vector-only memory layers treat every memory as a similarity hit; this collapses the distinction between "what should always be visible to the model" (persona, current goal, active constraints) and "what should be retrieved on demand" (historical episodes, archived knowledge). OS-style memory management formalizes the distinction. Core memory always lives in the prompt; recall and archival memory are fetched via tool calls. The model can rewrite its own core memory mid-conversation, mimicking working-memory consolidation.

Primary use cases

Long-lived personal-assistant agents that must remember preferences and ongoing projects across sessions; customer-support agents with multi-month conversation histories that exceed any context window; multi-agent systems where each agent has a distinct persona that must persist across handoffs; research agents accumulating archival domain knowledge over weeks.

Recent developments

Latest signals

Rebranded from MemGPT to Letta. The original MemGPT paper (Berkeley) → Letta SDK + Letta Cloud productization, but the OS-style memory primitives are preserved across the rebrand. Per Letta — repo.
The OS-style core/recall/archival split is influencing the broader field. Mem0 + Animesis CMA + EverMemOS all explicitly position against or extend the Letta primitive set. Per Mem0 — State of AI Agent Memory 2026.
Letta exposes MCP server endpoints. Letta memory blocks are reachable as MCP tools, letting any MCP-aware client (Claude Desktop, Cursor, etc.) consume Letta memory without the Letta SDK. Per the Letta repo.
Animesis CMA criticism: OS-memory ≠ ontological memory. The Animesis paper argues that Letta's OS-paging frames memory as a performance problem; CMA reframes it as an identity-continuity problem and treats Letta-style architectures as utility-tier, with constitutional governance as a strictly required layer above. Letta itself has not adopted CMA primitives yet. Per arXiv 2603.04740 — Animesis CMA.