Rather than storing all knowledge in model parameters or repeatedly retrieving raw text, we study memory as a reusable computational substrate for large language models and the agents built on them.
The project has two parts: native model memory, which looks at learned memory modules inside language models, and memory mechanisms for LLM agents, which looks at how agents keep and use what they learn over long, multi-step interactions.
Part 1: Native Model Memory
Learned external memory can be separated from the backbone that produced it. We study whether such memory can be frozen and reused by different target models, and how a model should decide when to rely on retrieved memory and when to generate memory instead.
Portable Memory
Detaching learned memory from its source model and reusing it across model families and tokenizers through lightweight, target-side readers.
Adaptive Memory
Routing between direct learned-memory retrieval and generated memories, so that generated memory acts as a selective correction rather than a replacement.
Publications
Frozen Memory Is Not Enough: Rethinking External Memory as Extraction
Mingyuan Li, Guangsheng Yu, Xu Wang, and Shaoxiong Ji
MemoryATHENA: Adaptive Routing over Latent and Generated Memories
Mingyuan Li, Guangsheng Yu, Juyuan Zhang, Xu Wang, Zhibo Man, Haonan Zhang, and Shaoxiong Ji
Part 2: Memory Mechanisms for LLM Agents
Agents that act over many steps and sessions need to keep track of past actions, context, and acquired knowledge beyond a single context window. We study mechanisms for what agents should remember, how that memory is organized and retrieved, and how it is updated as tasks and environments change.
Writing & Consolidation
Deciding what an agent should store from its interactions, and how to summarize and consolidate experience into lasting memory.
Retrieval & Context Management
Bringing the right memories into the context at the right time across long-horizon, multi-step tasks.
Updating & Forgetting
Keeping agent memory accurate and compact by revising outdated entries and discarding what is no longer useful.
Publications from this part are in progress.