The enterprise context layer stopped being a diagram this week. On October 23, Zoom acquired Onyx, an open-source platform for enterprise search and knowledge management founded in 2023. On October 20, Neo4j announced its Context Engine, a product that builds a reusable knowledge layer for agents. At Zoomtopia the same day, Zoom launched its own Unified Context Layer and an agentic search spanning more than 40 third-party applications. A suite vendor bought retrieval. A graph vendor shipped meaning. Both called it context, and the difference between the two is the work buyers now have to judge.
What Zoom actually bought
Onyx connects to more than 50 enterprise applications, covering documents, wikis, ticketing systems and customer records, and serves real-time context to people and to AI agents on demand, according to Zoom's announcement. Zoom put numbers on the traction: a community of more than 30,000 developers, customers including NASA, Nebius and Hudson River Trading, and more than one million user queries per week. Zoom did not disclose terms in its announcement. What it did disclose is a set of commitments: Onyx, led by co-founders Chris Weaver and Yuhong Sun, will remain a standalone platform and will stay open source, model-agnostic, app-agnostic and self-hostable inside Zoom AI Services. Zoom frames the deal inside its Conversation to Completion strategy, where a conversation is only as valuable as the decisions and actions that follow it.
Three days earlier, Zoom had already moved in the same direction without an acquisition. Its October 20 rollout of agentic search covered Zoom Meetings, Phone and Chat plus more than 40 third-party applications, including ServiceNow and Microsoft OneDrive, alongside the Unified Context Layer introduced at Zoomtopia. Onyx therefore does not start a strategy. It fills the retrieval layer of a strategy Zoom had announced before the deal closed.
What an enterprise context layer has to hold
Neo4j starts from the opposite end. Its Context Engine, now in private preview, builds what the company calls a knowledge layer: a reusable, shared substrate that agents use to reason consistently over enterprise knowledge. The launch rests on a simple observation. Data had two consumers: people, who understood relationships informally, and applications, which held them in hardcoded logic. Agents are a third consumer, and they inherit neither. A field that means suspended in one system means something else in another, and no amount of retrieval fixes a meaning the agent was never given.
The category now has an analyst definition. In its August 2026 Market Overview for AI Context Platforms, Gartner defines these platforms as software that lets data and AI teams construct, govern and deliver semantically organized knowledge to agents, treating context as a first-class enterprise asset rather than an agent-by-agent configuration and separating the life cycle of organizational knowledge from the life cycle of individual agents. Neo4j answers that definition with structure: a six-layer business ontology, a build and runtime layer over sources such as Databricks and Snowflake, a governance studio, and a graph-native memory that logs agent reasoning as decision tracking. Analyst Sanjeev Mohan, quoted in the launch, puts the distinction plainly: context is not just more data, it is about relationships, not just that two things are connected, but why and how.
Retrieval is not a record
The two launches agree on the diagnosis and diverge on the remedy, and the divergence is where procurement should press. Onyx serves context at query time and leaves the underlying systems as the record. Neo4j builds an explicit record of meaning and of past agent decisions, and leaves retrieval to GraphRAG over that record. Neither announcement answers the question yesterday's post on shared agent memory ended on: when the shared layer is wrong, who corrects it, how does the correction reach every agent that already read the wrong version, and what expires on its own? A context layer widens the blast radius of a stale fact in exactly the way a shared memory does, only across more applications.
There is a second test, and it applies to Zoom. A context layer that sits inside one suite has to stay neutral to earn trust outside that suite. Zoom's four pledges, standalone, open source, model-agnostic and self-hostable, are the right pledges, and they are checkable: the public repository and its licence are where any change would first appear. The pillar on self-editing context made the related point for working context: what an agent is allowed to keep, rewrite or discard decides what it can be trusted to do next. The same discipline applies one level up, to the layer every agent in the company reads.
That is the standard Jilu is built to. A context layer that agents can read but people cannot open, inspect and correct is a faster route to a confident wrong answer, not a company record. Retrieval decides what an agent sees. A record decides what an agent is entitled to believe: it keeps its sources, checks permissions at query time, and holds a review gate before a fact joins it. Zoom and Neo4j have now priced that distinction into their roadmaps, from opposite ends. Buyers should judge both launches, and every product in this category including this one, on the correction path, not the connector count. For the reasoning behind that position, see How Jilu Works.