September 2026 Summaries

Keenable integrates with the open-source agent memory platform cognee to provide more selective and current web content for knowledge graphs, helping avoid ingesting navigation, cookie banners, footers, and other irrelevant page material. Starting with cognee version 1.5, users can fetch a URL through Keenable and specify an extraction prompt of up to 2,000 characters to return only requested content, such as pricing tables or plan limits, before adding it to cognee for chunking and entity extraction. Users can also enable live fetching through a configuration flag or environment variable, ensuring pages are read directly rather than retrieved from Keenable’s indexed copy; scheduled live reads update a page node’s SHA-256 hash when its content changes. Setup requires installing cognee and setting a Keenable API key, while users can select Keenable explicitly through the fetch API, with further configuration details available in Keenable’s cognee integration documentation.

Keenable Search has integrated with Baseten Hosted Tools, allowing developers to provide models with live web search and page-fetch capabilities within a single inference request without managing separate API keys, billing, tool harnesses, or custom agent loops. Baseten executes the tools and returns retrieved content to the model, enabling it to search, read sources, fact-check, and continue reasoning during the same request. The partnership emphasizes that native web access should be evaluated by complete-task answer quality, end-to-end latency, and total cost, including tokens used to process retrieved content, rather than by individual search-call performance alone. Baseten is Keenable’s first inference-platform partner, and the companies plan to continue improving context management, search quality, speed, and affordability to make web access a standard capability across model experiences.

Keenable introduces SELECT, an agentic web-search interface that uses SQL-like queries to turn broad slices of live web content into structured, cited tables rather than returning conventional ranked links and snippets. The company argues that language models often produce typical answers to underspecified questions despite containing more specialized knowledge, and that access to external evidence can help agents identify less common but better-supported solutions. SELECT is designed for “wide search” tasks where answers depend on distributions, comparisons, or combinations of information across many pages, such as market pricing, rental trends, product launches, and multi-part costs. It applies LLM-powered semantic operators to extract, filter, aggregate, and join information from unstructured pages, following concepts from database research including Stanford’s LOTUS and commercial SQL features such as Snowflake Cortex AISQL. Keenable contrasts its declarative approach with standard search APIs and Perplexity’s Search as Code, claiming that agents should specify the information they need while the system handles collection and structuring. The company reports comparable benchmark performance to Perplexity’s approach on WideSearch and WANDR, while emphasizing that SELECT can be used with different models and agent frameworks.