Finds that filtering knowledge at 'write-time' (ingestion) maintains 100% RAG accuracy under noise levels where standard 'read-time' filtering completely collapses.
March 18, 2026
Original Paper
Selective Memory for Artificial Intelligence: Write-Time Gating with Hierarchical Archiving
arXiv · 2603.15994
The Takeaway
Suggests a fundamental shift in how we build memory systems for AI. Instead of retrieving from a noisy store, models should use salience-based gating during encoding, achieving 9x lower query-time costs and significantly higher robustness to distractors.
From the abstract
Retrieval-augmented generation stores all content indiscriminately, degrading accuracy as noise accumulates. Parametric approaches compress knowledge into weights, precluding selective updates. Neither mirrors biological memory, which gates encoding based on salience and archives rather than deletes superseded information. We introduce write-time gating that filters incoming knowledge objects using composite salience scores (source reputation, novelty, reliability) while maintaining version chai