# Big Updates to Redis: Context Engine, Tools, and Memory Simba Khadder, Redis | MLOps Community | 14:53 Source: https://www.youtube.com/watch?v=qlJg-TJFEnI Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks. Page: https://mlopstalks.com/talks/big-updates-to-redis-context-engine-tools-and-memory Published: 2025-12-16 Tags: context-engineering, mcp, memory, tool-use ## TL;DR - Redis is combining structured data, unstructured data, and memory into one context layer for agents. - Agents need semantic schemas that let them explore relationships and retrieve structured data instead of generating ad hoc SQL. - A unified MCP-native context engine lets agents choose among vector search, text search, geo indexes, key-value lookups, and memory. ## Summary Simba Khadder describes Redis's move from caching, rate limiting, and session storage toward providing real-time context for agents. He separates context into unstructured data, structured data, and memory. RAG works for one-shot questions over documents, but it does not reach the most valuable information in relational databases or live APIs. Khadder criticizes text-to-SQL because incorrect queries, weak access controls, and poor performance can affect production systems. Directly converting REST APIs into MCP tools also creates too many endpoints and parameters for an agent to use well. His proposed solution is a semantic graph or catalog that describes business entities, relationships, retrieval methods, and access rules. Redis's Context Engine brings these sources together behind an MCP-native interface so an agent can explore the data model, traverse relationships, search text, apply structured filters, retrieve memory, and call the right operations. ## Key ideas ### Redis is shifting from application speed to real-time agent context [02:37](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=157s) Khadder explains that Redis began with web caching, rate limiting, and session storage. In an agentic setting, he sees Redis providing real-time context to models. That context includes structured data exposed through MCP, unstructured data retrieved through RAG and a vector database, and memory. Redis remains backed by other data sources, but its role is the layer between those sources and the agent. ### Vector search handles document questions well, but it cannot reach every useful source [03:28](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=208s) Khadder describes RAG as processing documents, splitting them into chunks, and placing them in a search system so an agent can retrieve relevant context. Redis also supports text search, vector indexes, geo indexes, and hybrid queries. This helps tune retrieval from unstructured data. However, relational databases and live APIs sit outside the normal RAG workflow, even though they may contain the most valuable business information. ### Memory gives agents both conversation state and reusable user attributes [04:50](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=290s) Short-term memory keeps the state of an ongoing conversation, so a follow-up question can use what the user just said. Long-term memory stores attributes across conversations, such as user preferences. Khadder places both forms of memory alongside retrieval because they supply context that can be reused by the model for later requests. ### Structured business data is often the missing source for useful agent answers [05:50](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=350s) Khadder argues that valuable company data often lives in relational systems such as Postgres or Oracle. For an order-delay question, document retrieval might return generic reasons for delays. A useful answer needs the actual order, user, and item information from the database. This is the type of structured context that agents have often been unable to access directly. ### Text-to-SQL creates security and production risks [07:10](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=430s) Khadder calls ad hoc text-to-SQL dangerous in production. The generated SQL may be wrong, while row-level and column-level security rules are difficult to get right. A query can also create a performance problem on a production system and affect other users. He describes text-to-SQL as only a thin abstraction over raw SQL execution. ### Directly turning REST endpoints into MCP tools overwhelms agents [07:54](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=474s) Mapping every endpoint in an OpenAPI specification into an MCP tool can work for a small, carefully designed case. Enterprise APIs usually require chains of calls and expose many parameters that are irrelevant to an agent. Their shape often reflects how an organization is structured rather than the business task an agent needs to complete. Authentication is another problem that a simple converter does not solve. ### A semantic graph gives agents a model of the business and its allowed retrieval paths [09:21](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=561s) Khadder proposes writing a semantic graph in code. The catalog describes entities such as customers, orders, and items, along with their attributes and relationships. It can also control what an agent may see in different situations. Retrievers and the semantic catalog are combined into an MCP endpoint, giving the agent a structured way to understand and access the data. ### Agents can explore and traverse data instead of choosing from a fixed function list [10:24](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=624s) In Khadder's demo, Claude connects to an MCP endpoint and explores the data model before retrieving information. It lists orders, changes page sizes, gets the user linked to an order, and traverses the hierarchy to investigate fraud transactions. Khadder says this lets the agent do the reasoning while the semantic layer supplies the paths and data it can use. ### Redis wants one context surface for retrieval, tools, and memory [12:57](https://www.youtube.com/watch?v=qlJg-TJFEnI&t=777s) Khadder says structured data, unstructured data, and memory have historically been split across separate servers, databases, APIs, and frameworks. The Redis Context Engine puts them behind one semantic and access layer powered by MCP. An agent can choose among vector search, text search, geo indexes, key-value lookups, relationships, and memory while solving a task. ## Notable quotes - "The problem is no longer intelligence. The problem is getting the right context to the model to be able to let it think through it." (12:21) - "You need to describe. It's funny, we've talked about semantic data and semantic cataloges for a long time in the MLOps community, but the actual value of it has become way more even than it used to be because the semantic catalog here is not for people. It's for agents." (09:21) - "It's one surface of context that the agent can navigate and decide how it wants to navigate based on all the indices that we have from geoindex to vector index to search to just a key value lookup to get what it needs to solve problems." (14:09) - "A lot of people kind of start to use agents in ways where it's essentially a dramatized if statement like we're just using the agent and saying pick one of these four functions to run." (12:42) ## Tools & references mentioned - Redis - Featureform - enrich MCP - Model Context Protocol - MCP - Claude - RAG - Postgres - Oracle - OpenAPI ## Who should watch - You are building an agent that needs access to relational data, APIs, or user history in addition to documents. - Your team is considering text-to-SQL or an automatic OpenAPI-to-MCP conversion and needs to understand the operational risks. - You work on data access or agent infrastructure and want an example of using semantic schemas to guide retrieval and tool calls. ## Related talks - [Context Engineering 2.0: MCP, Agentic RAG & Memory](https://mlopstalks.com/talks/context-engineering-2-0-mcp-agentic-rag-memory) (Simba Khadder, Redis, 24:35) - [Turning Redis into a Composable, ML Data Platform](https://mlopstalks.com/talks/turning-redis-into-a-composable-ml-data-platform) (Samuel Partee, Redis, 48:21) - [Context Engineering 2.0](https://mlopstalks.com/talks/context-engineering-2-0) (Simba Khadder, Redis, 45:34) - [Overcoming Agentic Memory Management Challenges](https://mlopstalks.com/talks/overcoming-agentic-memory-management-challenges) (Biswaroop Bhattacharjee, Prem AI, 50:35) - [Dynamic Contextual Retrieval in Enterprise Analytics](https://mlopstalks.com/talks/dynamic-contextual-retrieval-in-enterprise-analytics) (Dirk Petzoldt, explai.com, 20:06)