TRACE: open-source hierarchical memory for LLM agents, 82.5% on MemoryAgentBench’s EventQA using gpt-oss-20B [P]
TRACE introduces a hierarchical memory system for LLM agents, organizing conversation history into topic trees instead of flat RAG chunks. It achieves 82.5% F1 on MemoryAgentBench’s EventQA task using the gpt-oss-20B model.
- TRACE organizes LLM agent memory into hierarchical topic trees instead of flat RAG chunks, improving retrieval accuracy.
- Achieves 82.5% F1 on MemoryAgentBench’s EventQA task using gpt-oss-20B, outperforming Mem0 (GPT-4o-mini).
- Released as an open-source PyPI package (`pip install trace-memory`) for easy integration.
- Demonstrates the potential of hierarchical memory structures for scalable, long-term agent memory management.
Researchers have unveiled TRACE, an open-source memory system designed to enhance LLM agents by structuring conversation history into hierarchical topic trees rather than relying on flat retrieval-augmented generation (RAG) chunks. This approach aims to improve the accuracy and efficiency of memory retrieval for agents handling complex, multi-topic dialogues.
The system was evaluated on MemoryAgentBench’s EventQA task, a benchmark introduced at ICLR 2026, which focuses on accurate event retrieval from agent conversations. TRACE achieved an F1 score of 82.5% using the gpt-oss-20B model, outperforming the baseline Mem0 (GPT-4o-mini) and demonstrating the potential of hierarchical memory structures in agentic workflows.
Available as a PyPI package, TRACE can be installed via `pip install trace-memory`, making it accessible for developers to integrate into their own LLM agent systems. The hierarchical design not only improves retrieval accuracy but also offers a scalable solution for managing long-term memory in dynamic environments.
Source: TRACE: open-source hierarchical memory for LLM agents, 82.5% on MemoryAgentBench’s EventQA using gpt-oss-20B [P]. Read the full piece at the source.
Provides an open-source tool to improve LLM agent memory retrieval with a hierarchical approach.
Enables more accurate and efficient agentic workflows, potentially reducing costs and improving user experience.
Offers a practical example of hierarchical memory systems in AI agents, useful for research and learning.
Highlights advancements in making AI agents more capable of handling complex, multi-topic conversations.
- RAG (Retrieval-Augmented Generation)
- A technique where an AI model retrieves relevant information from a database or knowledge base to improve its responses.
- F1 score
- A metric that balances precision and recall, used to evaluate the accuracy of a model's predictions.
- EventQA
- A benchmark task in MemoryAgentBench focused on accurate event retrieval from agent conversations.
Mantle cell lymphoma artificial intelligence prognostic index using hematoxylin and eosin histology - Nature
MIRA: Multiplayer Interactive World Models trained on Rocket League [R]
