“AI is not a buzzword anymore. It’s about which organization is implementing it faster, and at scale,” said Kanika Gupta, Solutions Architect at Redis, as she opened her deep-dive session at DevSparks Hyderabad 2025. For Kanika, the race is no longer about whether companies will adopt AI, but how quickly they can overcome its bottlenecks of speed, accuracy, and cost.
She reminded the audience that most developers know Redis as a caching solution, but that story is changing. “Redis is beyond caching,” she said. “We are the working memory layer for AI.”
Kanika explained that Redis is positioning itself as a backbone for next-generation applications from conversational agents and recommendation systems, to real-time analytics and fraud detection.
Tackling latency, cost, accuracy
One of the biggest challenges, Kanika noted, is latency.
“It sometimes takes five to ten seconds to generate an LLM answer; we don’t have that kind of time. Speed is something we cannot compromise on,” she explained.
The cost of repeatedly hitting large language models, combined with the risk of hallucinations and the complexity of integrating thousands of new AI tools, only compounds the problem.
Her session focused on how Redis is tackling these issues head-on. By serving as a vector database for retrieval-augmented generation (RAG), Redis can store embeddings and metadata together, enabling hybrid and full-text search at in-memory speeds. “Redis Enterprise serves as a fully featured vector database, requiring no extra setup or installation,” Kanika said, pointing out that Redis has been benchmarked as 62% faster than the second-best vector database.
Beyond retrieval: Semantic caching and agentic memory
Kanika also highlighted Redis’s role in solving one of AI’s costliest inefficiencies: redundant queries to LLMs. “Anytime I ask my payment app the same question someone else has asked, it shouldn’t have to hit the LLM again,” she explained. The solution is semantic caching, which Redis now offers as a service through LangCache, allowing developers to cache and retrieve responses for semantically similar queries.
The other breakthrough she described was Redis’s potential as a memory layer for agents. “If I ask a bot in the morning to book a flight, and in the evening I ask it to change that flight, it should have the context,” Kanika said. “That is where Redis comes into the picture…as the memory of the agent.”
Beyond caching and retrieval, Redis is broadening its role in data infrastructure. With Redis Data Integration, developers can keep source systems and caches in sync in real time, while RedisFlex enables tiered storage, keeping hot data in memory and moving warm or cold data to SSDs. Redis is also investing in ecosystem readiness, offering client libraries in multiple languages and integrations with popular AI frameworks such as OpenAI, LangChain, and AWS Bedrock.
Performance: The non-negotiable
Much of Kanika’s session circled back to a single priority — delivering performance at scale. Whether powering chatbots, fraud detection, or recommendation engines, Redis is positioning itself not as a supporting act but as a critical layer in the AI stack. “It is an in-memory vector DB plus accuracy, so speed and accuracy can be achieved together,” she said.
Her closing remarks captured the ambition Redis now carries into the AI era: “Redis is not just one caching solution. Start using it as your memory and intelligence layer for AI.”
