RAG Memory Configuration
Memory benchmarks (N=500, gpt-4o and gpt-5-mini readers routed per question type): 85.6% on LongMemEval-S at $0.0090 per correct answer (84.2% with gpt-4o reading every question), and 70.2% on LongMemEval-M, the 1.5M-token, 500-session variant. The LongMemEval paper (Wu et al., ICLR 2025) reports 65.7% (round, top-5), 71.4% (session, top-5) and 72.0% (round, top-10) on M. Benchmarks · Run JSONs · Write-up
HyDE Retrieval
HyDE improves RAG and memory retrieval by generating a hypothetical answer before
Hybrid BM25 + Dense Retriever
What this is
Session-Level Hierarchical Retriever
What this is
Citation Verification
Per-claim citation verification for AI agents: split an answer into claims, embed them with the retrieved sources, grade each claim supported, weak or unverifiable by cosine similarity, and mark contradictions with an optional NLI check.
Multimodal RAG (Image + Audio)
Memory benchmarks (full N=500, gpt-4o and gpt-5-mini readers routed per question type): 85.6% on LongMemEval-S at $0.0090 per correct; with gpt-4o reading every question, the same retrieval scores 84.2%, level with Mastra Observational Memory (84.23%). 70.2% on LongMemEval-M on the 1.5M-token / 500-session haystack variant — the only open-source library on the public record above 65% on M with publicly reproducible methodology. The same text-first retrieval pipeline that produced these numbers is what the multimodal pattern below indexes against (derived captions, transcripts, OCR, document text) once you have a text representation. Benchmarks · Run JSONs · SOTA writeup
Document Ingestion
Ingest files, folders and URLs (PDF, DOCX, HTML, Markdown, text, CSV, JSON, YAML) into a Memory brain as chunked, full-text-searchable traces.
Incremental Vector Ingestion
Keep a raw IVectorStore collection in sync with content-hash caching: skip unchanged chunks, re-embed only what changed, and re-run the same ingest as often as you like for near-zero cost. The lower-level path beneath Memory.ingest().
Library Index
A library of sources over any vector store. A source, such as a kept transcript or a document, is held as one unit: its passages are indexed under a tenant and access groups, filed in a folder, tagged and titled; it is replaced whole, re-scoped or removed; and a search reads only what its scope may see.