SandLogicTechnologies/Hy-Memory-30B-A3B-GGUF

Hy-Memory-30B-A3B

Hy-Memory-30B-A3B is a specialized long-term-memory writer model developed by Tencent Hunyuan for deciding what information should be retained in an external memory system and how that information should be represented. It is designed to operate as the memory-writing component rather than as a general-purpose answer-generation model.

The model uses a 30B-parameter Mixture-of-Experts architecture with approximately 3B active parameters per token. It is trained specifically for persistent-memory operations such as extracting useful information, updating existing memories when new evidence arrives, and resolving conflicts between previously stored and newly observed information.

The GGUF variants in this repository provide reduced-size representations of Hy-Memory-30B-A3B for local inference. These conversions make the memory-writing model more practical to deploy while preserving its underlying MoE architecture and specialized long-term-memory behavior.

---

Model Overview

• Model Name: Hy-Memory-30B-A3B
• Base Model: tencent/Hy-Memory-30B-A3B
• Parameter Count: Approximately 30 Billion Parameters
• Active Parameters: Approximately 3 Billion Parameters per Token
• Architecture: Mixture-of-Experts Language Model
• Modalities: Text
• Languages: English, Chinese
• Context Length: Up to approximately 262K tokens
• Primary Function: Long-Term Memory Writing
• Developer: Tencent Hunyuan
• License: Apache 2.0

Hy-Memory-30B-A3B is intended to function inside an external memory framework. It determines which information is worth retaining, how precisely that information should be encoded, and how newly observed evidence should modify the persistent memory bank rather than simply producing a conversational response.

---

Quantization Formats

This repository provides multiple GGUF quantized versions of Hy-Memory-30B-A3B designed to reduce the storage and memory requirements of the original F16 representation. The supplied formats provide different trade-offs between compression level and numerical precision.

IQ4_NL

• Size reduction of approx. 71.48% (15.98 GB) compared to 16-bit (56.03 GB)
• 4-bit non-linear quantization providing a substantially smaller representation of the 30B MoE model
• Retains more weight information than the supplied 3-bit variant while considerably reducing the model footprint
• Suitable for local memory-writing workloads where persistent-memory quality and deployment efficiency both matter
• Provides a practical middle ground for memory extraction, memory updates, and conflict-resolution workflows

IQ4_XS

• Size reduction of approx. 72.98% (15.14 GB) compared to 16-bit (56.03 GB)
• Compact 4-bit representation designed for deployments with tighter storage and memory constraints
• Produces a smaller file than IQ4_NL while remaining within the same general 4-bit precision class
• Suitable for local experimentation with long-term-memory writing and external memory management
• Provides an alternative when reducing model footprint is preferred while retaining a 4-bit representation

IQ3_M

• Size reduction of approx. 77.96% (12.35 GB) compared to 16-bit (56.03 GB)
• Aggressive 3-bit quantization providing the smallest GGUF representation among the supplied conversions
• Significantly lowers storage requirements for a large 30B-parameter MoE model
• Suitable for memory-constrained deployments and experimentation with specialized memory-writing workloads
• The lower numerical precision introduces a stronger compression trade-off for demanding memory extraction, evidence reconciliation, and long-context reasoning tasks

---

Training Background (Original Model)

Hy-Memory-30B-A3B is trained specifically for long-term-memory writing rather than conventional instruction following. Tencent's training approach places the model inside memory-framework configurations and optimizes its ability to construct useful persistent memory states from conversational or task-related information.

Memory On-Policy Distillation

• Uses Memory On-Policy Distillation (OPD) as the first stage of the training process
• Provides dense teacher supervision over memory-writing trajectories produced by the student model
• Trains the model to determine what information should be retained within a persistent memory bank
• Helps the model learn how memory states should evolve as additional information becomes available
• Establishes the initial memory-writing behavior before reinforcement learning

Attributed Memory Reinforcement Learning

• Uses attributed reinforcement learning as the second major training stage
• Evaluates memory updates according to their contribution to downstream memory performance
• Assigns credit at the memory-chunk level rather than treating an entire rollout as a single undifferentiated result
• Allows the model to learn which retained information improves future retrieval and answering
• Optimizes memory writing for long-horizon interactions and evolving memory states

External Memory Frameworks

• Designed to operate as the write component within frameworks such as Mem0, MemOS, and Hy-Memory
• Supports memory construction for conversational, working-memory, and coding scenarios
• Builds memory banks from an initially empty state and incrementally updates them with new information
• Evaluated across multiple memory harnesses using fixed retrieval and answer-generation configurations
• Demonstrates transfer to unseen memory frameworks such as EverMemOS according to the original evaluation setup

---

Key Capabilities

• Long-Term Memory Writing
  Identifies information that should be retained and converts relevant interaction content into persistent memory.

• Memory Extraction
  Extracts useful facts, preferences, events, and other information from incoming interactions for downstream storage.

• Memory Updating
  Revises existing memory when newly observed information provides additional or conflicting evidence.

• Conflict Reconciliation
  Helps determine how new evidence should affect previously stored information rather than blindly appending every new observation.

• Memory Compression and Representation
  Determines an appropriate level of detail for representing retained information inside a persistent memory bank.

• Long-Horizon Memory Management
  Maintains useful information across multiple interaction windows and evolving sequences of evidence.

• Cross-Framework Memory Writing
  Can serve as a specialized memory writer in external systems such as Mem0, MemOS, and Hy-Memory.

• MoE-Efficient Inference
  Uses approximately 3B active parameters per token despite having approximately 30B total parameters, allowing the sparse architecture to provide a larger parameter pool while activating only a subset for each token.

---

Usage Example

Using llama.cpp

The GGUF model can be loaded through llama.cpp for local inference. Since Hy-Memory is designed as a specialized memory-writing model, the generated output should be integrated into an external memory pipeline rather than treated as a conventional conversational answer.

./llama-cli \
  -m SandlogicTechnologies/Hy-Memory-30B-A3B-IQ4_NL.gguf \
  -p "Extract the information from the following interaction that should be retained in long-term memory and update the existing memory when new evidence conflicts with previously stored information."

---

Recommended Usecases

• Long-Term Memory Systems
  Build persistent-memory pipelines that retain important information across multiple conversations or task sessions.

• Conversational Memory
  Maintain useful user-specific information across long-running conversational interactions.

• Working Memory Management
  Construct and update temporary or persistent state used by an external agent during extended workflows.

• Coding-Agent Memory
  Preserve useful engineering knowledge, previous task information, and relevant historical context across coding sessions.

• Memory Conflict Resolution
  Handle situations where new observations contradict or modify information already present in the memory bank.

• External Memory Frameworks
  Integrate a specialized memory writer into systems such as Mem0, MemOS, and Hy-Memory.

• Long-Horizon Agent Research
  Experiment with persistent memory for agents operating over extended sequences of interactions and tasks.

• Local Memory Infrastructure
  Deploy the GGUF variants locally when the original F16 representation requires more storage or memory than the target environment provides.

Hy-Memory-30B-A3B is best suited to applications where the model acts as a memory-writing component. It should not be treated as a replacement for a general-purpose conversational LLM, since its primary optimization target is deciding what information should enter or modify an external persistent memory bank.

---

Acknowledgments

These quantized models are based on the original work by the Tencent Hunyuan development team.

Special thanks to:

• The Tencent Hunyuan team for developing and releasing Hy-Memory-30B-A3B and its long-term-memory research framework.

• The research team behind the Hy-Memory framework for developing memory-writing methodologies based on Memory On-Policy Distillation and attributed reinforcement learning.

• The open-source memory-system community for advancing research into persistent memory, long-horizon agents, and external memory architectures.

• Georgi Gerganov and the llama.cpp open-source community for enabling efficient GGUF conversion and local model inference.

---

Contact

For questions, feedback, or support, please reach out at support@sandlogic.com or visit https://www.sandlogic.com/.