Robot Memory System

Your Robot Forgets Everything. Until Now.

The memory system that robots have been missing. Remember skills, learn from failures, build on experience —

Get Started → How It Works
pip install robotmem
Open Source (Apache 2.0)
9
Integrations
∞
Cross-Session Recall
< 10ms
Recall Latency
100%
Offline Capable
Features

Built for Real Robots

Not another chatbot memory. RobotMem is designed for physical agents that interact with the real world.

Multi-Modal Storage

Store all 5 perception types — visual, tactile, auditory, proprioceptive, and procedural — with numeric parameters and trajectories.

Cross-Session Learning

Robot remembers across reboots. Skills learned today are available tomorrow. No retraining needed.

Model Agnostic

Works with any framework — ROS, Isaac Gym, MuJoCo, dm_control. Not locked to any specific model or vendor.

Semantic Search

Find relevant experiences by meaning, not keywords. "How did I grasp the red cup?" returns the right memory.

Auto Deduplication

dHash for visual similarity, Jaccard for text overlap. Keeps memory clean without manual curation.

Trajectory & Parameters

Store force profiles, joint trajectories, and numeric parameters as structured data alongside natural language.

Framework Integrations

Plug Into Your Stack 9 integrations

Drop-in adapters for the tools you already use.

Simulation

Isaac Lab Persistent memory for GPU-accelerated training ManiSkill Experience replay across ManiSkill episodes robosuite Store manipulation experiences in robosuite LIBERO Transfer learned skills across LIBERO tasks Gymnasium-Robotics Spatial memory for GoalEnv environments

Algorithms

stable-baselines3 Memory callback for SB3 training loops LeRobot Sync robot experiences to HuggingFace Hub

Deployment

ROS 2 Native ROS 2 memory service node Unitree RL Gym Persistent memory for Unitree robot training
How It Works

Three Steps to Robot Memory

Simple API. No infrastructure. Just save experiences and recall them when needed.

1

Experience

Robot performs an action. RobotMem captures the perception — what it saw, felt, and did — with full context.

2

Remember

Next session starts. RobotMem retrieves relevant past experiences via semantic search. Robot picks up where it left off.

3

Evolve

Over time, patterns emerge. Skills crystallize. The robot builds genuine expertise from accumulated experience.

Quick Start

Robot Memory in 13 Lines

Python API & MCP Server. Your robot remembers in minutes.

robot_controller.py 
from robotmem import save_perception, recall

# Save a grasping experience
save_perception(
    description="Grasped red cup: force=12.5N, 30 steps",
    perception_type="procedural",
    data='{"actions": [[0.1, -0.3, 0.05]], "force_peak": 12.5}',
)

# Next session: recall similar experiences
memories = recall("how to grasp a cup")
for m in memories["memories"]:
    print(m["content"], m["_rrf_score"])
Terminal Output 
$ python robot_controller.py
[recall] hybrid mode | 3 results | top score: 0.847
Grasped red cup: force=12.5N, 30 steps    score=0.847
FetchPush: success, dist 0.012m, 28 steps score=0.723
Push cube: force=11.8N, overshoot 0.03m   score=0.651
Why RobotMem

Built Different

Existing AI memory systems are designed for chatbots. RobotMem is designed for robots.

Capability RobotMem Mem0 Zep Letta
Target Use CaseRobotsChatbotsChatbotsChatbots
Multi-modal perception✓ 5 types✕✕✕
Trajectory storage✓✕✕✕
Numeric parameters✓✕✕✕
Model agnostic✓✓✕✕
Offline capable✓ Local ONNX✕ Cloud✕ Cloud✓
Visual dedup (dHash)✓✕✕✕
MCP protocol✓✕✕✕
Natural language storage✓✓✓✓
Score8 / 82 / 81 / 82 / 8
Proven Results

Real Experiments, Real Numbers

Tested in MuJoCo robotics environments. Not theoretical — measured.

+25%
FetchPush Success Rate
Baseline 42% → with memory 67%. Same policy, same environment — the only difference is persistent memory.
3×
Cross-Environment Transfer
Train in FetchPush, transfer to FetchSlide: baseline 4% → with memory 12%. No model fine-tuning — just memory.
< 10ms
Recall Latency
Hybrid BM25 + vector search over 10,000 memories. Fast enough for real-time robot control loops.
0
Cloud Dependencies
100% offline. Local SQLite + ONNX embeddings. No API keys, no internet, no vendor lock-in.

Start Building Robots That Remember

Open source. Free forever. One pip install away.

pip install robotmem
Star on GitHub Read the Docs