<p>Volatile memory contains valuable forensic evidence, but retrieving relevant strings at forensic scale is difficult because memory-extracted corpora are large, noisy, fragmented, multilingual, and often relevant without sharing exact lexical forms. Conventional keyword and regular-expression workflows remain effective when the expected wording is known, but they are limited for synonymic, cross-lingual, or corrupted evidence. This paper investigates a hybrid semantic–lexical retrieval strategy for scalable evidence discovery over unstructured memory strings. The study uses a two-phase design. Phase I introduces a controlled benchmark spanning 10 forensic concepts, 10 retrieval categories, and four languages to evaluate exact, semantic, cross-lingual, and corruption-tolerant retrieval. Dense multilingual embedding models perform best for semantic and cross-lingual retrieval, whereas fuzzy lexical matching is strongest for corrupted-string recovery. Phase II transfers the learned thresholds into a forensic-scale corpus of 4.45 million memory-derived strings. The strongest dense models reduce the candidate review space by 98.66–99.24% while preserving substantial recall, and fuzzy matching remains uniquely effective for corrupted-memory artifacts. Additional validation analyses show that these conclusions remain stable under threshold perturbation and preprocessing audits. A GPU-based embedding stage provides an approximately 6 × speedup over CPU, approximate nearest-neighbor indexing reduces dense-retrieval latency by approximately 43 × with near-equivalent target recovery, and a two-node sharded exact-search prototype achieves a 1<i>.</i>93 × speedup while preserving exact output. Together, the results show that memory-string forensics is best supported by a complementary retrieval workflow: exact search when terms are known, dense retrieval for semantic and multilingual triage, and fuzzy lexical matching for corrupted-memory recovery.</p>