GLACIER is an open-source project by Dopove Private Limited that gives State Space Models like Mamba a persistent, time-aware memory. It solves context rot by integrating a lightweight virtual memory engine (ICE-Lite) and a temporal reranking layer (Temporal-RAG), ensuring models retain long-term memory while remaining fast and efficient.
This project was architected and built by Saran S, Founder & CEO of Dopove.
- GitHub: https://github.com/Saran-386
- Dopove: https://www.dopove.com/
For a fast setup on Linux with CUDA, you can install the pre-built glacier-ice-lite wheel directly:
pip install https://github.com/Dopove/Glacier/releases/download/v0.1.0/glacier_ice_lite-0.1.0-py3-none-any.whl- Persistent Memory: Stop and restart sessions with full memory recall.
- Temporal-Aware RAG: A time-decay and validity-aware retrieval layer prevents the model from using stale information.
- Agentic Capabilities: Native support for multi-step tool use, "Turbo-Stitching" for long-form generation, and file ingestion.
-
$O(1)$ Inference Speed: Retains the core speed advantage of Mamba by managing all memory externally.
GLACIER's architecture is significantly more efficient than traditional Transformer models for long conversations. By avoiding a full-context KV-cache, it maintains constant speed and a minimal token footprint.
Explore the core principles that make GLACIER superior to standard RAG and memory wrappers:
- Visual Architecture Overview: Mamba's External Hippocampus.
- Defeating Context Rot: How we solve SSM amnesia.
- Multi-Tiered Memory: The 100 Billion Token Horizon.
- GLACIER vs. The Competition: Why our architecture wins against Mem0, Zep, and LangGraph.
To guarantee a perfectly configured environment with all CUDA dependencies, we strongly recommend using Docker.
# 1. Build the Docker image (this will take 10-15 minutes)
docker build -t glacier .
# 2. Run the full benchmark suite inside the container
docker run --gpus all glacierThis is the fastest and most reliable way to run GLACIER.
Click here for local installation instructions
Warning: Installing mamba_ssm from source is complex and requires a specific CUDA and C++ compiler environment. This method is only recommended for advanced users.
Prerequisites:
- NVIDIA GPU with CUDA Toolkit 12.1 installed.
- A compatible C++ compiler (e.g.,
gcc-11orgcc-12).
# 1. Create and activate a virtual environment
python3 -m venv .venv && source .venv/bin/activate
# 2. Install dependencies (this will compile CUDA extensions)
pip install -r requirements.txt
# 3. Download embedding models (if not already present)
mkdir -p ~/.cache/ice/models/
wget https://huggingface.co/Xenova/all-MiniLM-L6-v2/resolve/main/onnx/model_quantized.onnx -O ~/.cache/ice/models/model.onnx
wget https://huggingface.co/Xenova/all-MiniLM-L6-v2/resolve/main/tokenizer.json -O ~/.cache/ice/models/tokenizer.json
# 4. Run the full benchmark suite
python tests/test_glacier.pyYou can deploy the interactive Gradio demo (examples/app.py) to HuggingFace Spaces in minutes. See the Getting Started guide for more details.
/
├── src/ice_lite/ # The core ICE-Lite Python package
├── vendor/mamba_ssm/ # Mamba2 source (from state-spaces/mamba)
├── temporal_rag.py # The Temporal-RAG source (by Emmimal P Alexander)
├── tests/ # Test suites
│ ├── test_glacier.py # The full GLACIER benchmark test suite
│ └── test_agentic_features.py # Tests for tool use, stitching, and ingestion
├── examples/ # Demos and examples
│ └── app.py # Gradio demo for HuggingFace Spaces
├── docs/ # All project documentation
├── setup.py # Packaging script for the `glacier` package
├── requirements.txt
└── LICENSE # Main project license (Apache 2.0)
We welcome contributions! Please see our Contributing Guide for details on how to get started.
This project is licensed under the Apache License 2.0. See the LICENSE and NOTICE files for full details, including third-party component licenses.