Awesome-Memory-for-Robotics

🤖 Memory is a fundamental capability to store, retrieve, and utilize past experiences for improved perception, planning, and decision-making. This mechanism enables robots to learn from experience, adapt to new situations, and operate effectively in complex, dynamic environments.

Here we organized a list of resources on memory for robotics, including papers, code, datasets, and benchmarks.

Scope: Robotics + Memory , not generic LLM memory.

---

📋 Table of Contents

• Surveys & Overviews
• Memory Systems & Cognitive Architecture
  • Contextual Memory
  • Episodic Memory
  • Semantic Memory
  • Long-term Memory
• Memory for Manipulation
  • Robotics Foundation Models (RFMs)
  • Mobile Manipulation
  • General Manipulation
• Memory for Navigation
• Memory for Locomotion
• Spatial & Scene Memory
  • 3D Scene Memory
  • Spatial Memory & SLAM
• World Models & Memory
• Lifelong Learning & Continual Learning
  • Experience Replay
  • Skill Transfer & Memory
• Benchmarks & Evaluation
• Datasets
• Contributing

---

Taxonomy

Memory in robotics often mixes multiple axes:

• Timescale
  • Contextual memory (within-episode, short-term working memory)
  • Episodic memory (experience replay / trajectories / videos)
  • Semantic memory (facts, object affordances, task graphs)
  • Long-term memory (stable knowledge across tasks)
• Representation
  • key-value / slot memory, RNN / LSTM / Transformer state
  • retrieval over embeddings (RAG / vector DB)
  • 3D maps / NeRF / Gaussian splats / voxel grids
  • symbolic graphs (scene graphs / task graphs)
• Usage
  • policy conditioning (retrieve then act)
  • planning (search over remembered states)
  • perception (temporal integration / object permanence)
  • safety & failure recovery (remember mistakes)

Read more: docs/taxonomy.md

---

Surveys & Overviews

There hasn't been a comprehensive survey on memory for robotics, we hope this list can help.

---

Memory Systems & Cognitive Architecture

People built memory for robots from our cognitive science. Here are some classical papers:

---

Further, I divide memory into four categories: Contextual memory, episodic memory, semantic memory, and long-horizon memory.

Contextual Memory

Working memory maintains task-relevant information during execution. It enables robots to track object states, maintain spatial awareness, and hold intermediate computation results. It is usually implemented by keyframe, history and context-window.

---

Episodic Memory

Episodic memory stores specific events and experiences from the robot's past, enabling recall of particular situations, trajectories, and outcomes. It is usually implemented by memory banks, replay buffers, or latent tokens.

---

Semantic Memory

Language provides robots with structured knowledge about concepts, objects, and their relationships. It is usually implemented by Neural Symbolic Model, Chain of Thought Reasoning, or Scene Graphs.

---

Long-term Memory

Long-term memory enables continuous accumulation of knowledge across tasks and sessions, supporting lifelong learning and adaptation without catastrophic forgetting. It is usually implemented by experience replay, skill transfer, or continual learning.

Memory for Manipulation

Robotics Foundation Models (RFMs)

Because non-markovian property makes RFMs like VLAs prone to overfit, it's a huge challenge for RFMs to learn memory. We need to design spatial-temporal memory mechanisms carefully.

Mobile Manipulation

General Manipulation

---

Memory for Navigation

Memory is of great importance for navigation, there has been rich research on it.

---

Memory for Locomotion

---

Spatial & Scene Memory

3D Scene Memory

Representing and reasoning about 3D environments.

Spatial Memory & SLAM

Simultaneous Localization and Mapping(SLAM) is the most fundamental explicit type of memory.

---

World Models & Memory

World models that incorporate memory for prediction and planning.

---

Lifelong Learning & Continual Learning

Experience Replay

Lifelong Learning Frameworks

Skill Transfer & Memory

---

Benchmarks & Evaluation

We treasure these features that actually pose challenges to memory:

• Partial observability: occlusions, hidden states, deferred rewards
• Long-horizon: requires recalling earlier observations/instructions
• Distribution shift: new scenes, objects, layouts
• Recovery: can the robots use memory to diagnose and retry?

Read more: docs/evaluation.md

---

Datasets

See: docs/datasets.md

---

Citation

If you find this repository useful, please consider citing:

@misc{awesome-memory-for-robotics,
  author = {Jie Wang},
  title = {Awesome Memory for Robotics},
  year = {2025},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/Everloom-129/Awesome-Memory-for-Robotics}}
}

---

Contributing

PRs welcome! Please read the contribution guidelines.

To add a new paper or resource:

1. Fork this repository
2. Add your entry in the appropriate section
3. Submit a pull request

Please ensure your PR follows the existing format and includes:

• Paper title
• Venue and year
• Links to paper, code, and project page (if available)

---

Related Awesome Lists

• Awesome Memory for Agents — [repo]
• Awesome Multimodal Memory — [repo]
• Awesome LLM Robotics — [repo]

---

License

This work is licensed under CC0 1.0 Universal.

---

⭐ Star this repository if you find it useful! ⭐