Unified Memory Agent design
According to the model card, UMA incrementally maintains a compact core summary and a structured key-value Memory Bank, with one policy handling memory construction and question answering.
Open Source Model Profile · ICTNLP
UMA-4B
UMA-4B is a 4.41B-parameter Qwen3-family text-generation model from ICTNLP. According to the model card, it is the Generalist checkpoint of the Unified Memory Agent for long-context reasoning.
- Publisher
- ICTNLP
- Task
- text-generation
- Model type
- qwen3
- License
- apache-2.0
- Library
- transformers
- Publication status
- Approved for indexing
Model overview
UMA-4B is published by ICTNLP as a Qwen3 text-generation model with memory-agent, tool-use, and long-context tags. The captured configuration identifies Qwen3ForCausalLM and Safetensors metadata reports 4,411,424,256 parameters. According to the model card, the base model is Qwen/Qwen3-4B-Instruct-2507 with BF16 Safetensors weights.
Recorded capabilities
GRPO training method
According to the model card, training used end-to-end reinforcement learning with Task-Stratified GRPO.
Generalist checkpoint role
According to the model card, this repository holds the Generalist checkpoint for Test-Time Learning and Accurate Retrieval evaluations, distinct from the Ledger-QA specialist checkpoint.
vLLM serving example
According to the model card, the model can be served with vLLM using max-model-len 16384 and gpu-memory-utilization 0.8, alongside a Transformers loading example.
Use cases in the source record
- Long-context reasoning experiments using the tool-using memory-agent design with a core summary and key-value Memory Bank.
- Memory construction and question-answering workflows through explicit memory and retrieval operations with the official repository runners.
Limitations and unknowns
- According to the model card, agent behavior depends on prompt templates, tool implementations, retrieval backend, chunking policy, and inference configuration.
- According to the model card, the model was primarily trained and evaluated on English-language research benchmarks.
- No independent evaluation results were extracted from this record.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: ICTNLP/UMA-4B
Captured: Unknown. Processed: 2026-09-07T19:35:39.839412+00:00.
UMA-4B (Generalist) UMA-4B is the Generalist checkpoint of the Unified Memory Agent (UMA) introduced in Learning to Remember: End-to-End Training of Memory Agents for Long-Context Reasoning . UMA is a tool-using memory agent that incrementally maintains a compact core summary and a structured key-value Memory Bank. The same policy performs memory construction and downstream question answering through explicit memory and retrieval operations. Code: github.com/ictnlp/unified-memory-agent Paper: arXiv:2602.18493 Specialist checkpoint: ICTNLP/UMA-LedgerQA-4B Checkpoint Variant This repository contains the Generalist UMA checkpoint used…
F001F002F003F004F005F006F007F008F009F010F011F012F013F014F015F016F017F018