ESC
↑↓ navigate ↵ open esc close
Getting Started › Introduction

Introduction

robotmem is a persistent memory system for robotic AI agents. It lets robots remember skills, learn from failures, and build on past experience — across sessions, across reboots, across lifetimes.

Why robotmem?

Robots run thousands of experiments, but each episode starts from zero. robotmem stores every experience — parameters, trajectories, successes and failures — and retrieves the most relevant ones to guide future decisions.

Core Capabilities

6 API Tools

Tool Purpose
learn Record physical experience (parameters, strategies, lessons)
recall Retrieve experience — BM25 + vector hybrid search with context_filter and spatial_sort
save_perception Store perception/trajectory/force data (visual / tactile / proprioceptive / auditory / procedural)
forget / update Delete or correct erroneous memories
start_session / end_session Episode lifecycle (automatic consolidation + proactive recall)

5 Perception Types

robotmem understands 5 types of robotic perception:

Structured Experience Retrieval

Not just vector search — robotmem understands the structure of robotic experience:

# Only retrieve successful experiences
recall(query="push to target", context_filter='{"task.success": true}')

# Find nearest spatial scenario
recall(query="grasp object", spatial_sort='{"field": "spatial.object_position", "target": [1.3, 0.7, 0.42]}')

# Combined: success + distance < 0.05m
recall(
    query="push",
    context_filter='{"task.success": true, "params.final_distance.value": {"$lt": 0.05}}',
)

Context JSON — 4 Partitions

Every memory's context is structured into 4 semantic partitions:

{
    "params":  {"grip_force": {"value": 12.5, "unit": "N", "type": "scalar"}},
    "spatial": {"object_position": [1.3, 0.7, 0.42], "target_position": [1.25, 0.6, 0.42]},
    "robot":   {"id": "fetch-001", "type": "Fetch", "dof": 7},
    "task":    {"name": "push_to_target", "success": true, "steps": 38}
}

recall automatically extracts params / spatial / robot / task as top-level fields in every returned memory.

Architecture at a Glance

SQLite + FTS5 + vec0
├── BM25 full-text search (jieba CJK tokenization)
├── Vector search (FastEmbed ONNX, pure CPU)
├── RRF fusion ranking
├── Structured filtering (context_filter)
└── Spatial nearest-neighbor sorting (spatial_sort)

Comparison

Dimension MemoryVLA (Academic) Mem0 (Product) robotmem
Target users Specific VLA model Text AI Robotic AI
Memory format Vectors (unreadable) Text Natural language + perception + params
Structured filter Not supported Not supported Supported (context_filter)
Spatial retrieval Not supported Not supported Supported (spatial_sort)
Physical params Not supported Not supported Supported (params partition)
Installation Paper code compile pip install pip install
Database Embedded Cloud service Local SQLite

What's Next