Remnic vs Letta.
This is the least head-to-head comparison on the site. Letta is a runtime you build agents in. Remnic is a memory layer under agents that already exist. The real question is which layer you need.
What Letta is
Letta, formerly MemGPT, treats the LLM like an operating system: the agent manages its own memory hierarchy, paging context in and out and editing its own memory as it works. It is a full open-source agent runtime, architecturally closer to LangGraph than to a drop-in memory layer. If you are building a new stateful agent from scratch, it is a serious option.
Where Letta wins
- You control the whole agent loop, and memory management is native to it.
- Self-editing memory is a research direction with real momentum, and Letta is its reference implementation.
- One framework covers agent orchestration and memory together.
Where Remnic differs
- No new runtime. Remnic sits beneath Claude Code, Codex CLI, Cursor, Pi, Hermes, OpenClaw, and any MCP client through hooks and MCP tools. Your agents stay exactly what they are; they just stop forgetting.
- Human-governed writes over self-editing. Remnic's bet is that memory an agent edits on its own authority is memory you cannot trust later. Writes are importance-gated, corrections flow through a contract, and contested pairs wait for your resolution.
- Provenance and measurement. Recall X-ray explains every retrieved memory, and MemCorrect measures the correction path. See Recall X-ray and benchmarks.
- Files you own. Plain markdown on your disk, MIT licensed, free.
Switching
There is no automated Letta importer today; Letta memory lives inside its runtime state. If you are moving agents off a Letta deployment, ask on the issue tracker and describe your export shape. For everything else, Remnic starts learning from your first session.