RobotMem vs Zep
Which memory system should you use in 2026?
Quick Summary
Zep excels at one thing: managing conversation memory for AI assistants and chatbots. It tracks dialogue history across sessions, extracts user facts ("user prefers metric units"), and retrieves relevant context to improve chatbot coherence over time. It is cloud-first and commercially supported. RobotMem operates in an entirely different domain — it gives physical robots the ability to store and recall sensory episodes: what the robot saw, felt, heard, and how it moved its joints. It works fully offline on edge hardware and exposes a Model Context Protocol server. The two tools do not compete; they serve different systems. Zep is for words; RobotMem is for embodied experience.
Bottom line: Choose RobotMem for robots, autonomous vehicles, or any embodied system that produces sensory data. Choose Zep for conversational AI products that need persistent user context and managed cloud infrastructure.
Feature Comparison
| Capability | RobotMem | Zep |
|---|---|---|
| Target use case | ✓ Physical robots | Chatbots & AI assistants |
| Multi-modal perception | ✓ 5 types (visual, tactile, auditory, proprioceptive, procedural) | ✕ Text / dialogue only |
| Trajectory storage | ✓ Joint trajectories + timestamps | ✕ Not supported |
| Numeric parameters | ✓ Force, velocity, torque arrays | ✕ Not supported |
| Offline / edge capable | ✓ Local ONNX embedding, no internet required | ✕ Cloud-first, requires hosted service |
| Visual deduplication | ✓ dHash perceptual hashing | ✕ Not supported |
| MCP protocol | ✓ Built-in MCP server | ✕ Not supported |
| Natural language storage | ✓ Supported | ✓ Core feature |
| Model agnostic | ✓ ROS, MuJoCo, Isaac Gym, dm_control | ✕ Tied to Zep cloud infrastructure |
| Dialogue history management | ✕ Not the focus | ✓ Core strength |
| Community & ecosystem | ~ Early stage, growing | ✓ Established community, LangChain integration |
| Managed cloud option | ✕ Self-hosted only | ✓ Commercial cloud product |
| License | Apache 2.0 | Apache 2.0 (community) / commercial |
| Installation | pip install robotmem |
pip install zep-python + hosted service |
Detailed Comparison
1. Memory Model: Sensory Episodes vs. Conversation Sessions
Zep's architecture revolves around sessions — sequences of user/assistant message turns. Each session accumulates facts about the user and dialogue context, which Zep surfaces in future turns. This is the right model for a chatbot that needs to remember a user's name across months of conversations. RobotMem's memory model is fundamentally different: it stores sensory episodes — timestamped bundles of perception data (RGB images, force readings, joint angles) linked to an outcome. A robot does not have conversations; it has experiences. RobotMem stores those experiences in the form they were produced, not as serialized text.
2. Offline vs. Cloud-First Architecture
Zep is designed as a hosted service. Even the open-source community edition requires running a Zep server with a Postgres backend, and the commercial product assumes cloud connectivity. For robots deployed in warehouses, agricultural fields, or industrial facilities — environments where internet access is intermittent or absent — this is a fundamental constraint. RobotMem is built edge-first: the embedding model runs locally via ONNX Runtime, the database is local SQLite or Postgres, and no external API calls are made at inference time.
3. MCP Protocol: Querying Robot Memory from Agent Loops
RobotMem's built-in Model Context Protocol server lets any MCP-aware agent (Claude, local LLMs via Ollama, LangChain agents) call memory as a structured tool. This is increasingly important as multi-agent robotics systems use LLMs for high-level planning. Zep does not offer an MCP interface, so wiring Zep into an MCP-native agent pipeline requires custom glue code.
4. Where Zep Wins: Conversational AI Depth
For its intended use case, Zep delivers features RobotMem has no interest in: temporal knowledge graphs that track how user facts change over time, automated session summarization, fact extraction from dialogue, and a commercial SLA. If you are building a customer support bot, a personal AI assistant, or any product where long-term conversational coherence matters, Zep offers proven infrastructure that RobotMem was never designed to provide.
Frequently Asked Questions
What is the main difference between RobotMem and Zep?
RobotMem stores multi-modal sensory data — images, joint trajectories, force readings, audio — produced by physical robots during operation. Zep stores conversational history and user facts for chatbots. The data models are designed for completely different domains: embodied sensory experience vs. language-based interaction history.
Can I use Zep for robot memory?
Zep can store text, so you could serialize robot logs as strings, but you lose trajectory structure, numeric precision, multi-modal typing, and offline capability. Zep is also cloud-hosted, which conflicts with the offline requirements of most physical robot deployments. The integration work required would exceed the benefit.
Is RobotMem free to use?
Yes. RobotMem is open source under Apache 2.0 and free forever — install with pip install robotmem. Zep has a free community edition, but its commercial product with full features and cloud hosting is a paid service.
Ready to give your robot persistent memory?
Open source, fully offline, one pip install away.
Get Started on GitHub Read the Docs