Cross-Source Fusion Search for Heterogeneous AI Memory Systems

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: Jia, Charles
Natura: Recurso digital
Pubblicazione: Zenodo 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901623563878400
author Jia, Charles
author_facet Jia, Charles
contents <p><br>  AI coding assistants maintain two distinct types of memory: semantic memory (code<br>  indexes) and episodic memory (learned facts and decisions). These memories typically<br>  reside in separate databases with different schemas, lifecycles, and write patterns.<br>  Existing hybrid search approaches (RRF, DBSF, linear combination) assume a single data<br>  source, while multi-memory systems (Mem0, MemGPT, Zep) either delegate source selection<br>  to the LLM or return unranked concatenations.</p> <p>  We present a cross-source fusion algorithm that exploits a shared embedding space to<br>  produce unified rankings across heterogeneous databases. Our multiplicative scoring<br>  formula addresses a previously undocumented vote asymmetry problem in additive<br>  multi-source fusion, where independent per-source BM25 rankings inflate text-match<br>  influence relative to a unified vector signal.</p> <p>  Key contributions:<br>  - Three-level search architecture separating per-source and cross-source fusion<br>  - Multiplicative fusion formula eliminating vote asymmetry in multi-source BM25<br>  - Graceful degradation to BM25-only mode when vector search is unavailable<br>  - 7x latency reduction via parallel retrieval (620ms → 86ms)</p> <p>  The approach requires no changes to existing per-source search pipelines.</p> <p>  Project: https://github.com/gladego/index1</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18556780
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Cross-Source Fusion Search for Heterogeneous AI Memory Systems
Jia, Charles
hybrid search
rank fusion
<p><br>  AI coding assistants maintain two distinct types of memory: semantic memory (code<br>  indexes) and episodic memory (learned facts and decisions). These memories typically<br>  reside in separate databases with different schemas, lifecycles, and write patterns.<br>  Existing hybrid search approaches (RRF, DBSF, linear combination) assume a single data<br>  source, while multi-memory systems (Mem0, MemGPT, Zep) either delegate source selection<br>  to the LLM or return unranked concatenations.</p> <p>  We present a cross-source fusion algorithm that exploits a shared embedding space to<br>  produce unified rankings across heterogeneous databases. Our multiplicative scoring<br>  formula addresses a previously undocumented vote asymmetry problem in additive<br>  multi-source fusion, where independent per-source BM25 rankings inflate text-match<br>  influence relative to a unified vector signal.</p> <p>  Key contributions:<br>  - Three-level search architecture separating per-source and cross-source fusion<br>  - Multiplicative fusion formula eliminating vote asymmetry in multi-source BM25<br>  - Graceful degradation to BM25-only mode when vector search is unavailable<br>  - 7x latency reduction via parallel retrieval (620ms → 86ms)</p> <p>  The approach requires no changes to existing per-source search pipelines.</p> <p>  Project: https://github.com/gladego/index1</p>
title Cross-Source Fusion Search for Heterogeneous AI Memory Systems
topic hybrid search
rank fusion
url https://doi.org/10.5281/zenodo.18556780