RobotMem vs Letta (MemGPT)

Which memory system should you use in 2026?

Quick Summary

Letta (formerly MemGPT) is a stateful agent framework where an LLM manages its own memory autonomously — deciding when to move information from in-context working memory to archival storage, and when to retrieve it. This is a powerful paradigm for long-horizon text-based tasks like research assistants or coding agents. RobotMem takes a different approach: it uses a deterministic, typed storage engine to persist structured sensory data from physical robots — RGB images, joint trajectories, force-torque arrays, audio clips — without relying on an LLM to make storage decisions. Where Letta is intelligent about managing text, RobotMem is efficient about managing sensor streams. Both are offline-capable, both are open source, but they serve fundamentally different systems.

Feature Comparison

Detailed Comparison

1. LLM-Managed vs. Deterministic Memory

Letta's defining insight is that an LLM can manage its own memory — the model itself decides to call archival_memory_insert or archival_memory_search based on what it thinks is important. This allows for flexible, intelligent memory management in open-ended agent tasks. The trade-off is that every memory operation requires an LLM inference step, adding latency and cost. RobotMem takes the opposite approach: storage is deterministic — when a sensory episode arrives, it is stored immediately without LLM involvement. Search uses vector similarity over pre-computed ONNX embeddings. For robots processing hundreds of frames per second, the LLM-in-the-loop model would be a bottleneck.

2. Data Types: Sensory Episodes vs. Archival Text

Letta's archival memory is optimized for text: documents, notes, conversation summaries. RobotMem's storage model was designed from the ground up for the data robots actually produce: typed perception records with fields for image tensors, proprioceptive joint vectors, force-torque readings, and audio waveforms. It also stores trajectory replays — sequences of joint angles with timestamps — which are essential for imitation learning and motion primitives. No text serialization of this data would preserve the fidelity needed for replaying or analyzing robot behavior.

3. Offline Capability: ONNX vs. Local LLM

Both tools can operate offline, but with different complexity profiles. Letta requires a full LLM backend — typically 7B-70B parameters via Ollama or similar — to function, since the LLM is core to memory management, not just retrieval. On an NVIDIA Jetson or Raspberry Pi 5, running a 7B parameter model alongside the robot control loop is often infeasible. RobotMem's local inference uses a compact ONNX embedding model that runs comfortably on edge hardware, with no LLM inference required during normal operation.

4. Where Letta Wins: Long-Horizon Agent Reasoning

Letta's architecture shines for tasks where the agent itself must reason about what is worth remembering over long time horizons — research assistants that read papers over weeks, coding agents that learn project conventions over months, or personal assistants that track preferences across years. The LLM-driven memory model is uniquely suited to these open-ended, text-centric tasks. RobotMem stores whatever the robot perceives without judgment; Letta's agent decides what matters. For high-level cognitive agents running on cloud compute with reliable internet, Letta is the more research-forward choice.

Frequently Asked Questions