Memory Retrieval — References

Retrieval mechanics that apply when the population is the agent's own store, and the evidence that what is retrieved is not necessarily what gets used.

Combining relevance with recency and use

• Generative Agents: Interactive Simulacra of Human Behavior — Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang and Michael S. Bernstein, ACM UIST,
  2023. A retrieval function over an agent's own memory stream combining recency, importance, and relevance, and the clearest published statement of why similarity alone is insufficient here. Cited by name.
• Reflections of the Environment in Memory — John R. Anderson and Lael J. Schooler, Psychological Science, 1991. The argument that the probability an item will be needed is predictable from how recently and how often it has been needed before, which is the basis for any decay term in a memory ranking. Cited by name.

Index and ranking mechanics

• Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs — Yury A. Malkov and Dmitry A. Yashunin, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. The index behind most vector search, including why a highly selective filter interacts badly with graph traversal. Cited by name.
• Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods — Gordon V. Cormack, Charles L. A. Clarke and Stefan Buettcher, ACM SIGIR, 2009. A simple, parameter-light way to combine rankings from independent retrieval methods. Cited by name.
• MTEB: Massive Text Embedding Benchmark — Niklas Muennighoff, Nouamane Tazi, Loïc Magne and Nils Reimers, EACL, 2023. A standard comparison across embedding models and tasks, useful as a shortlist and explicitly not as a substitute for evaluating on your own memory items. Cited by name.

What the model does with what is retrieved

• Lost in the Middle: How Language Models Use Long Contexts — Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang, Transactions of the Association for Computational Linguistics, 2024 (arXiv:2307.03172). Evidence that retrieval success and use are different things, and that position within the assembled input affects which retrieved items are actually used. Cited by name.

No verified reference is recorded for score thresholds over agent memory stores, for per-type decay rates, or for labelling conventions that let a model discount a stale recollection. These are engineering practice without published study.