Redis constructed its identify because the caching layer that stored internet functions from collapsing below load. The issue it’s concentrating on now has the identical construction however is more durable to resolve: manufacturing AI brokers failing not as a result of the fashions are flawed, however as a result of the info beneath them is scattered, stale and structured for people reasonably than machines. Retrieval pipelines constructed for single queries can’t take in the quantity brokers generate.
The hole Redis is concentrating on is structural: brokers make orders of magnitude extra information requests than human customers, however most retrieval layers have been constructed for the human-scale downside. Redis Iris, launched Monday, is the corporate's reply: a context and reminiscence platform that sits between an agent and the info it must act. The platform combines real-time information ingestion, a semantic interface that auto-generates MCP instruments from enterprise information fashions, and an agent reminiscence server constructed on Redis Flex, a rewritten storage engine that runs 99% of knowledge on flash at a tenth of the price of in-memory storage alone.
The announcement lands as enterprise RAG infrastructure is in energetic transition. VentureBeat's Q1 2026 VB Pulse RAG Infrastructure Market Tracker discovered purchaser intent to undertake hybrid retrieval tripling from 10.3% to 33.3% between January and March. Retrieval optimization surpassed analysis as the highest enterprise funding precedence for the primary time. Customized in-house retrieval stacks rose from 24.1% to 35.6% as enterprises outgrew off-the-shelf choices. Redis is just not the one infrastructure vendor studying these indicators — a number of information platform suppliers have repositioned round agent context layers in latest weeks.
The dimensions mismatch is the structural argument behind the launch.
"Corporations could have orders of magnitude extra brokers than human beings," Rowan Trollope, CEO of Redis, informed VentureBeat. "Orders of magnitude extra brokers than human beings means orders of magnitude extra load on again finish programs."
From cache to context
Trollope traces the parallel again to the cellular period: When legacy backends constructed for department tellers abruptly needed to serve one million smartphone customers, Redis grew to become the caching layer that absorbed the load with no full rebuild.
What’s totally different this time is that brokers can’t write their very own middleware. Within the cellular period, a developer would sit with a database administrator, determine the queries an software wanted and hard-code the caching logic right into a middleware layer. Brokers can’t do this. They should discover the appropriate information at runtime, by means of interfaces constructed for them prematurely, or they stall.
"That is just like the analogy of the grocery retailer within the fridge," he mentioned. "If each time it’s important to go make your sandwich, it’s important to run to the grocery retailer to get the meals, that's not very environment friendly. You set a fridge in each home, you retailer a bit of little bit of meals there. And that's sort of the place we nonetheless are likely to exist within the infrastructure stack."
What Redis Iris consists of
Iris ships 5 parts that collectively cowl information ingestion, semantic entry, reminiscence and caching.
Redis Information Integration. Now basically availability. RDI makes use of change information seize pipelines to sync information from relational databases, warehouses and doc shops into Redis repeatedly, with connectors for Oracle, Snowflake, Databricks and Postgres.
Context Retriever. Now in preview. Builders outline a semantic mannequin of enterprise information utilizing pydantic fashions and Redis auto-generates MCP instruments brokers use to question it instantly, with row-level entry controls enforced server-side. Trollope describes the shift from basic RAG as a directional inversion. "It's only a flip to let the agent pull the info as an alternative of presupposing and stuffing it into the pipeline," he mentioned.
Agent Reminiscence. Now in preview. Shops quick and long-term state throughout classes so brokers carry context with out re-deriving it on every flip.
Redis Flex. A rewritten storage engine that runs 99% of knowledge on SSDs and 1% in RAM, delivering petabyte-scale retrieval at sub-millisecond latencies.
Redis Search and LangCache. The retrieval and semantic caching spine beneath the platform. LangCache reduces redundant mannequin calls by caching immediate responses.
What analysts say
The info business is mostly heading in the identical path now. Each main database vendor is making a context layer argument.
Conventional database distributors including Oracle are integrating context and reminiscence layers to deliver relational databases into the agentic AI period. Function-built vector database distributors together with Pinecone are doing the identical, constructing out a brand new data layer for agentic AI context. Standalone context layers like Hindsight are additionally a part of the rising panorama.
Trollope frames Redis's place as structurally totally different from that competitors.
"For us to win, nobody else has to lose," he mentioned. Many Redis deployments already run MongoDB or Oracle because the backend system of file. Iris displays and caches from these programs reasonably than displacing them. Redis is launching Iris within the Snowflake market with native connectors.
Stephanie Walter, Apply Chief for AI Stack at HyperFRAME Analysis, places the market context plainly. "The market is converging on the identical conclusion: brokers don't simply want extra tokens or higher fashions. They want ruled, present, low-latency context," Walter mentioned.
Her learn on Redis's differentiation focuses on the place Redis already sits within the stack, which is near runtime, latency-sensitive operational state, and real-time information.,
"The pitch is just not 'higher RAG' as a lot as 'brokers want stay context, reminiscence, and quick retrieval whereas they’re really working," she mentioned.
Whether or not it's Redis or one other vendor, each context layer expertise will face a governance problem to achieve success.
"Agentic AI is not going to scale within the enterprise if each agent turns into a brand new value heart, a brand new information entry threat, and a brand new governance exception," she mentioned. "The profitable context layers would be the ones that make brokers sooner, cheaper, and safer to run."
For real-time scientific AI, getting context flawed is just not an choice
Mangoes.ai is one firm that has already needed to reply these questions in manufacturing, below circumstances the place the price of getting context flawed is measured in affected person outcomes.
Amit Lamba, founder and CEO of Mangoes.ai, runs a real-time voice AI platform deployed throughout giant healthcare services the place sufferers and clinicians ask stay questions on therapy, scheduling and case historical past. Mangoes.ai constructed its stack natively on Redis from the beginning.
"Retrieval, reminiscence, and session state all run by means of Redis, so we're not stitching collectively separate instruments and hoping they speak to one another," Lamba mentioned.
The issue Iris's dynamic reminiscence functionality addresses is what occurs throughout a fancy session.
"Take into consideration a one-hour group remedy session," Lamba mentioned. "It is advisable to know who mentioned what, when, and have the ability to floor the appropriate data to the therapist within the second. That's not a easy retrieval downside."
The platform runs a number of specialised brokers in parallel, one for entity identification, one for relationship reasoning and one for integrating case historical past.
"The dynamic reminiscence functionality maps virtually completely to the issue we're fixing," Lamba mentioned.
What this implies for enterprises
For enterprises that constructed their AI stack round RAG, the retrieval layer that bought them to manufacturing is not sufficient to maintain them there
The RAG period is giving approach to context structure. The basic RAG mannequin pushed information into the agent earlier than the mannequin was known as. Manufacturing deployments are flipping that: brokers pull what they want at runtime by means of software calls, treating the info layer as a stay useful resource reasonably than a pre-loaded payload. Groups nonetheless optimizing RAG pipelines are fixing final yr's downside.
The semantic layer is now manufacturing infrastructure. The mannequin that defines enterprise entities, their relationships and the entry guidelines between them must be constructed, versioned and maintained with the identical self-discipline as a knowledge pipeline. Most organizations haven’t staffed or structured for that work. The enterprises that outline their context structure now are those that won’t should rebuild it when agent workloads scale.
Price range is already transferring. VB Pulse Q1 2026 information exhibits retrieval optimization funding rising from 19% to twenty-eight.9% throughout the quarter, overtaking analysis spending for the primary time. Organizations that spent the earlier yr measuring their retrieval high quality are actually spending to repair it. The context layer is an energetic procurement determination, not a roadmap merchandise.
"The primary purchaser query shouldn’t be 'Do I want a vector database, lengthy context, reminiscence, or a context engine?' It ought to be 'What does this agent must know, how recent should that data be, who’s allowed to entry it, and what does each retrieval value?'" Walter mentioned.