An Index That Rewrites Its Own Keys When Retrieval Misses
Self-Evolving Search Index proposes SELF-INDEX, and the premise is one of those things that is obvious once somebody says it. Effective index representations vary across retrieval environments, and right now the person who tunes them for your corpus is a human who does it once and never again. Paper is arXiv 2609.19656 by Sangam Lee and colleagues, at https://arxiv.org/abs/2609.19656.
Two pieces do the work. An Optimizer watches for retrieval gaps and revises the index keys to close them. A Query Simulator goes looking for demand that has not shown up yet, generating plausible queries against the corpus rather than waiting for real ones to fail. Revisions get validated before they land, so the index is not free to drift on the basis of one bad episode. Reported gains are consistent across different corpora and different retrievers, which for an infrastructure paper matters more than a big number on one setup.
The reason this belongs in an agent conversation rather than a search conversation is the downstream claim. The authors report improvements for search agents and for retrieval in agent memory systems, and agent memory is exactly where static indexing hurts most. A memory store gets written by the agent, in the agent's own vocabulary, describing situations nobody anticipated when the schema was drawn. The gap between how something was filed and how it later gets asked for is the entire failure mode of agent memory.
Fair warning, the authors label it work in progress and it sits in cs.IR, so there is no agent benchmark headline number to quote.
Still, the direction is right and underexplored. Everybody is building agents that write to memory. Almost nobody is building memory that notices it is being written wrong.
← Back to all articles
Two pieces do the work. An Optimizer watches for retrieval gaps and revises the index keys to close them. A Query Simulator goes looking for demand that has not shown up yet, generating plausible queries against the corpus rather than waiting for real ones to fail. Revisions get validated before they land, so the index is not free to drift on the basis of one bad episode. Reported gains are consistent across different corpora and different retrievers, which for an infrastructure paper matters more than a big number on one setup.
The reason this belongs in an agent conversation rather than a search conversation is the downstream claim. The authors report improvements for search agents and for retrieval in agent memory systems, and agent memory is exactly where static indexing hurts most. A memory store gets written by the agent, in the agent's own vocabulary, describing situations nobody anticipated when the schema was drawn. The gap between how something was filed and how it later gets asked for is the entire failure mode of agent memory.
Fair warning, the authors label it work in progress and it sits in cs.IR, so there is no agent benchmark headline number to quote.
Still, the direction is right and underexplored. Everybody is building agents that write to memory. Almost nobody is building memory that notices it is being written wrong.
Comments