September 2026 Summaries

Engram is a managed memory and context service for AI agents that processes submitted conversations asynchronously through extraction, transformation, and commitment steps, with topic descriptions serving as prompts that determine which information becomes memory and in what form. Its default personalization template separates durable user facts, profile information, and optional conversation summaries, while carefully written exclusions can prevent transient incidents from being retained. Topic configuration also controls scope, partitioning, and whether a topic is bounded to a single continually updated memory, allowing applications to maintain singular profiles or per-conversation summaries while preserving separate user data. Memories can be retrieved through relevance-based vector, BM25, or hybrid search, direct topic fetching, or lookup by ID, and retrieval limits should be tuned to avoid adding irrelevant context. Prompt placement materially affects LLM caching costs: stable, always-needed memories can be fetched once near the start of a session, while changing search results should be appended after the current user message with an appropriate cache breakpoint, preserving cached system instructions and conversation history.

Weaviate 1.39 expands its Rotational Quantization (RQ) support with a new 4-bit option and broad performance improvements to 8-bit and 1-bit RQ, including SIMD-accelerated rotations, encoding, distance kernels, and CPU prefetching. The release substantially speeds encoding and distance calculations, improves import throughput, and introduces 4-bit RQ variants that reduce heap usage by about 45% relative to 8-bit while often delivering comparable or better recall-versus-query-performance results through lower memory-bandwidth demands. Optional mean centering and exact storage of two extreme rotated coordinates improve 4-bit recall, particularly for embeddings with nonzero means or anisotropic distributions, while rescoring candidates using original vectors can further approach exact-search accuracy. Tests from 1 million to 250 million vectors found RQ recall remained broadly stable, though ANN index settings may still require sharding at larger scales. Comparisons with TurboQuant found that centered RQ4 generally matched or outperformed enhanced TurboQuant configurations, while RQ also supports efficient 8-bit operation. RQ4 is available through the existing Weaviate RQ configuration using a 4-bit setting, with centering enabled as an optional feature, whereas 8-bit RQ remains the default in Weaviate Cloud.

HFresh is Weaviate’s disk-based vector index for applications that need to reduce memory consumption and support large, mutable datasets while accepting lower query throughput than in-memory HNSW. It partitions vectors into disk-resident postings, uses a compact HNSW centroid index in memory to identify relevant postings, and limits disk reads through a two-stage search that rescoring candidates with original vectors preserves ranking quality. HFresh applies RQ8 quantization to centroids and RQ1 compression to postings, substantially reducing memory and storage requirements, while asynchronous split, merge, and reassignment operations continuously rebalance partitions without full rebuilds. It also supports filtered search through bitmap allow lists, switching to direct exact search for highly selective filters and using posting-aware routing for broader filters. Benchmarks on a one-million-vector dataset showed HFresh using 239 MB of Go heap versus 6.67 GB for uncompressed HNSW, though HNSW retained higher throughput at similar recall; a separate test successfully built an HFresh index for one billion 256-dimensional vectors. Introduced as a technical preview in Weaviate 1.36 and generally available in 1.38, HFresh is aimed at workloads where memory efficiency and scale outweigh the need for the lowest latency.

Foundry is a prototype workflow for converting an existing, inconsistently organized creative archive into a searchable library while leaving all original files untouched. It performs a read-only scan to create a manifest containing metadata such as paths, file types, modification dates, sizes, and content hashes, then enriches records with descriptions, tags, extracted content, relationship roles, and source URIs that link search results back to original assets. These prepared records are synchronized with a Weaviate collection, where embeddings and metadata support keyword, semantic, and hybrid search without creating duplicates during repeated updates. Users can search by remembered concepts rather than filenames, preview relevant assets, and filter results by project, asset type, or role, with potential future filters for rights and approval information. The project demonstrates a path from local folders to retrieval across images, videos, documents, and other creative materials, while identifying automatic enrichment, incremental scanning, rights-aware filtering, and relevance feedback as planned improvements.

Late-interaction multi-vector retrieval offers an alternative to conventional PDF RAG pipelines that rely on OCR, text chunking, and single-vector text embeddings, which can miss information contained in charts, tables, layouts, and diagrams. Using Weaviate’s hosted multi2multivec-weaviate model, each PDF page is rendered as an image and represented by multiple vectors, allowing MaxSim-based matching between query concepts and relevant visual regions without text extraction or chunking. A demonstration using NVIDIA’s FY2026 quarterly investor presentations shows that the system can retrieve specific pages containing revenue and gross-margin trends even when the query wording does not appear on those pages. Weaviate Cloud supports drag-and-drop ingestion and can pair retrieval with its Query Agent to generate synthesized, page-cited answers, while an equivalent production workflow can be implemented in Python using PyMuPDF to rasterize pages and Weaviate to store and vectorize them. The approach is most suitable for visually rich, chart-heavy PDF collections, whereas conventional text embeddings may remain less expensive and adequate for text-dominant corpora; hybrid systems can combine both methods.