April 2026 Summaries

Weaviate v1.37 introduces new features and enhancements designed to extend its functionality and integration capabilities, particularly with AI agents and large-scale data operations. Key updates include the preview release of a built-in Model Context Protocol (MCP) server, enabling AI agents and IDEs to interact natively with the database, and Extensible Tokenizers offering accent folding and custom stopword presets for improved text analysis. Diversity Search with Maximum Marginal Relevance (MMR) provides query-time reranking to enhance result diversity, while Query Profiling offers detailed timing breakdowns for performance analysis. Incremental Backups streamline data management by storing references to unchanged files, reducing storage needs and speeding up the backup process. Additionally, support for audio as a modality in the multi2vec-google module and the BlobHash property type for efficient media vectorization are introduced. These features are available open-source and on Weaviate Cloud, with detailed guidance for those upgrading self-hosted versions.

Engram is a managed memory service designed to enhance agentic applications by integrating memory into systems, ensuring personalization, and enabling continuous learning. Built on the Weaviate vector database, Engram uses asynchronous pipelines to process raw data into structured memories, which are then stored and queried efficiently. The system addresses the challenges faced by long context windows in language models by maintaining memory as structured data rather than an ever-growing context. Engram allows for customization through templates and configurable pipelines, enabling users to define topics and scopes that determine how memories are extracted, categorized, and stored. The service supports various input types and includes features like buffers for batch processing, ensuring low latency and efficient memory management. Engram's flexibility makes it suitable for a range of applications, from chatbots to multi-agent systems, by providing a robust framework for memory management, allowing agents to learn and adapt based on historical interactions.

Weaviate Shared Cloud is now generally available on AWS in the US East and Europe, offering an AI-native platform that supports building AI applications with ease, regardless of whether infrastructure is on GCP or AWS. Users can select their preferred cloud provider and region when creating a cluster, benefiting from fully managed services, automatic upgrades, and robust security features such as granular RBAC and immutable backups. The platform includes a native embedding service for faster data vectorization, tools for creating RAG pipelines and agentic workflows, and a comprehensive console for data management and monitoring, eliminating the need for complex database queries. Weaviate Cloud is designed for teams that prefer AWS, have specific compliance needs, or want to transition from open source to a managed environment with extensive built-in tooling. Future plans include expanding to more regions and providers and introducing new features to enhance performance and usability.

Engram, a memory product built on Weaviate's vector search technology, is designed to enhance the workflow of a product manager by structuring memory around semantic categories, thus providing context that a traditional memory system like MEMORY.md cannot. The author, who integrated Engram into their workflow with Claude Code, observed that while Engram offers significant advantages in decision archaeology by recalling reasoning chains and document intent more efficiently, it encountered challenges in planning sessions where previous contexts were ignored. Engram's integration revealed silent failures in tasks perceived as forward-looking and added noticeable overhead to sessions. The solution involved developing deterministic retrieval triggers to ensure relevant memories are automatically injected at specific points, improving session fluidity without relying on Claude's discretion. Despite the initial expectation of proving Engram's value, the exercise highlighted specific conditions and mechanics that influence its efficacy, providing valuable insights for improvement as it moves towards general availability.

Multimodal embeddings offer a transformative approach to data retrieval by allowing searches across various formats like text, images, audio, and video within a unified embedding space, thus overcoming the traditional limitations of converting all data into text. This approach leverages contrastive learning, where paired data inputs from different modalities are trained to align closely in a high-dimensional space, enabling semantic search that captures the full spectrum of information without losing context or detail. Recent advancements in multimodal models, such as Google's Gemini Embedding 2, facilitate this by preserving important aspects of each data type, making it possible to effectively search and retrieve content based on meaning rather than format. Examples include querying audio files without transcripts, reading PDFs as complex visual documents, and finding specific moments in videos through semantic content retrieval. While multimodal embeddings are not a one-size-fits-all solution, they are particularly beneficial when dealing with data that contains non-textual signals, offering a more accurate and comprehensive retrieval experience.