Reproduce “Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon 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
Seyed Moein Abtahi, Rasa Rahnema, Hetkumar Patel, Neel Patel, Majid Fekri, Tara Khani
Memanto introduces a vector-only architecture for agentic memory, leveraging an information-theoretic search engine and a typed schema to outperform complex hybrid systems. This framework achieved 89.8% and 87.1% accuracy on the LONGMEMEVAL and LOCOMO benchmarks, respectively, while eliminating data ingestion costs and significantly reducing operational overhead compared to graph-augmented approaches.
Key insights
- High-fidelity agent memory can be achieved with a simpler vector-only architecture, challenging the prevailing assumption that complex knowledge graph components are essential.
- For agentic memory systems, maximizing recall (providing a broader context to the LLM) is more effective than aggressive precision filtering at the retrieval layer, as LLMs robustly filter irrelevant information.
- An information-theoretic search engine provides deterministic, exact-match semantic retrieval, enhancing the reproducibility and stability of agent behavior compared to approximate nearest neighbor methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.