Reproduce “HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinchang Zhu, Jindong Li, Cheng Zhang, Jiahong Liu, Menglin Yang
HeLa-Mem stores conversation turns as a graph and retrieves memories using both semantic similarity and learned links strengthened when memories are co-activated. On LoCoMo conversations averaging about 300 turns, it tests four question categories. On GPT-4o-mini, its multi-hop F1 is 40.14% versus MemoryOS† at 38.39%, but this result does not establish gains in every comparison or long-run scalability.
Key insights
- Dynamic graph-based memory with Hebbian learning allows for the progressive strengthening of associations between co-activated memories, mimicking biological memory.
- A dual-level memory architecture combining an episodic graph with a semantic store improves both fine-grained recall and the consolidation of reusable, structured knowledge.
- Mechanisms like spreading activation and active consolidation (through a Reflective Agent) are crucial for efficient retrieval of multi-hop information and managing memory growth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.