compare
dig vs Letta (MemGPT) — a memory layer, not an agent framework
Letta (formerly MemGPT) is an agent framework with tiered self-managed memory. dig is the data + memory layer underneath any agent — local, reversible, framework-agnostic.
Letta and dig both care about agent memory, but one is the agent and the other is the layer it stands on.
What each one is
Letta (formerly MemGPT) is an agent framework. It runs an agent that pages its own memory between tiers autonomously — a full runtime for building stateful agents.
dig is not an agent and not a model. It is the local, reversible data + memory layer an agent reads and writes: search a knowledge base, recall budgeted context, retain a note or a finished session, all journaled. You bring the agent; dig holds the knowledge.
Where they differ
- Layer. Letta is the agent. dig stays the backend — drive it from Letta, Claude Code, Codex, or your own loop over MCP.
- Lock-in. Letta is a framework you build inside. dig is harness-agnostic: one MCP server and a portable skill plug into anything, so you are never married to one agent runtime.
- Local + reversible. dig runs fully on your machine with no model required, and every change is undoable. Memory is auditable data, not opaque agent state.
- Files, not just memory. dig also organizes, dedupes, and reconciles the real files behind the memory to a policy you set.
When to use which
Use Letta if you want a batteries-included agent runtime that manages its own memory.
Use dig if you want a portable memory/retrieval backend that any agent or framework — Letta included — can use as its store, locally and reversibly.