Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Lingua: | |
| Pubblicazione: |
Zenodo
2026
|
| Soggetti: | |
| Accesso online: | https://doi.org/10.5281/zenodo.19788364 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
Sommario:
- This paper presents a three-tier memory architecture that mimics human selective recall: always-loaded base knowledge for sub-millisecond responses, a hierarchical concept index for O(1) concept-to-location lookup, and bounded working memory with LRU eviction. Unlike approaches that load entire knowledge graphs at startup, this architecture maintains constant startup time (less than 100ms) and bounded memory (less than 25MB) while preserving 95% reasoning accuracy. Validation across 9 agents in 5 domains demonstrates 97,000x query speedup over full-graph loading for large corpora and 96% tier-prediction accuracy.