archon-memory-core vs Letta

Letta (formerly MemGPT) is a full agent runtime. archon-memory-core is a memory library. Most teams evaluating both are really asking "do I want a framework that includes memory, or memory I can plug into my own framework?"

What each one actually is

Letta is an agent runtime that descends from the MemGPT paper. It ships as a Docker-deployed server with PostgreSQL, a REST API, a Python SDK, and a UI. Memory is a first-class concept: core memory blocks edited by the agent, archival memory backed by a vector store, and a recall mechanism triggered by context pressure. You write agents against Letta's abstractions.

archon-memory-core is a Python library. pip install archon-memory-core, import the client, call write() and retrieve(). It runs in-process against SQLite or your existing Postgres. There is no server, no runtime, no agent abstraction — just memory primitives with explicit persistence classes and contradiction resolution at retrieval.

Feature comparison

When to pick Letta

• You don't have an agent framework yet and want one with opinionated defaults.
• You want an out-of-the-box multi-agent server with a UI and REST endpoints.
• Your team is comfortable operating a Postgres-backed service.

When to pick archon-memory-core

• You already have a framework (LangGraph, CrewAI, Pydantic AI, something custom).
• You want contradiction resolution as a retrieval-time property, not an agent instruction.
• You don't want to run another server.
• You care about the AMB v2.3 preregistered result (99.2% top-1 on contradiction retrieval).

Using them together

Letta's core memory is great for the agent's working context (persona, current goals). archon-memory-core is great for the long tail of facts that need to survive, evolve, and sometimes contradict. Wire archon-memory-core as a tool Letta can call — keep Letta's runtime, add explicit contradiction handling underneath.

FAQ