Redis

Context Gateway gives an agent one Model Context Protocol endpoint for tools from any number of upstream sources: third party or in house MCP servers that speak Streamable HTTP, and Context Surfaces that generate query tools from a data model of records in Redis. The gateway holds each upstream's credentials in Vault, AWS Secrets Manager or encrypted Redis, so the agent never carries them, and an agent with access to a gateway can call every tool it exposes, including tools that change data in the connected system.

Sourceredis.io/docs/latest/develop/ai/context-engine/context-gateway/conceptsread 2026-11-09

Routing rather than execution is what the Iris suite offers multiagent systems. Intent Router matches each incoming message against example phrasings by vector search, with no model call, and returns the closest intents with scores and a target naming the agent, tool or workflow that should handle it, after which the customer's application dispatches. Agent Mesh, a beta Python SDK, keeps A2A task state in Redis so another process can take over a task, but no hosted runtime plans or runs multistep work.

Sourceredis.io/docs/latest/develop/ai/context-engine/intent-routerread 2026-11-09

Agent Playbooks keep a maintained store of the answers and procedures a business has approved: FAQs that return a full answer, and Agent Skills that return a preview before the agent loads the instructions and files. Entries carry embeddings, attribute filters and pins, are matched to the agent's task by semantic similarity, and keep immutable versions with publish, withdraw and rollback. Underneath sits Redis vector search with HNSW and flat indexes, hybrid metadata filters and the RedisVL library for retrieval pipelines.

Sourceredis.io/docs/latest/develop/ai/context-engine/agent-playbooks/conceptsread 2026-11-09

Agent generated knowledge passes a person before agents can use it. Background extraction turns resolved conversations into proposed FAQs that wait in a review inbox, and a reviewer accepts each one into an unpublished entry and then puts it live, with write, publish and propose granted to different keys. Ingestion webhooks on self managed deployments let a signed callback accept, reject or transform session events and generated memory changes before they are saved, and detectors redact or drop card numbers and other identifiers. Context Gateway has no approval step on tool calls.

Sourceredis.io/docs/latest/develop/ai/context-engine/agent-playbooks/conceptsread 2026-11-09

Console sign in supports SAML single sign on, started by either the identity provider or Redis Cloud, with just in time user creation and no SCIM, and team members hold one of six roles from Owner to Logs viewer. Each Iris service issues its own API keys with scoped grants, such as read, write, publish and propose on a playbook, and self managed suites mint, rotate and revoke agent keys through an Identity Service. The Redis trust center lists SOC 2 Type 2, ISO 27001, 27017, 27018 and 42001, PCI DSS and HIPAA.

Sourcetrust.redis.ioread 2026-11-09

Lens, the tracing service in Redis Iris, records every turn an agent handles: each retrieval and memory recall, what entered the prompt, every model call and tool call with its timing, and how the turn ended, grouped into sessions by agent and searchable in the console or over a read API with scoped keys.

Lens is in beta and open to any Redis Cloud account with an Essentials or Pro database, or self managed on Kubernetes. Iris services add their own steps without extra code, and traces flow into existing OpenTelemetry tools. Redis Cloud also keeps an account log of activity by originator, database and API key.

Sourceredis.io/docs/latest/develop/ai/context-engine/lensread 2026-11-09

Two memory tiers sit in a store the customer owns. Session memory keeps ordered events with actor, role and timestamp under a session ID, with a time to live of one day by default and up to 365 days, and summarizes older events automatically. Long term memory is extracted in the background, embedded for semantic search and scoped by owner, session, namespace, topic or memory type, with its own time to live of 365 days by default. Memories can be edited, deleted by ID, or removed for one user at once.

Sourceredis.io/docs/latest/operate/iris/agent-memory/create-serviceread 2026-11-09

Each Iris service stores its data in Redis databases the customer owns: on Redis Cloud, a database in the customer's account, placed in the AWS, Google Cloud or Azure region picked when it is created, and when self managed, the customer's own Redis. Every service also runs on the customer's Kubernetes cluster through one Helm release under a suite license, with the Identity Service included, and the Agent Mesh SDK runs inside the customer's application. Agent Memory on Redis Cloud cannot use Active Active, Flex or PrivateLink databases.

Sourceredis.io/docs/latest/operate/irisread 2026-11-09

The agent builder in the Redis developer guides generates working Python or JavaScript code for four starter agents: a recommendation engine, a conversational assistant, a knowledge assistant with semantic caching, and a conversational assistant on Iris Agent Memory, using OpenAI, Anthropic Claude or Llama 2. Tutorials walk through agents built with LangGraph, Google ADK and Flowise. Inside Iris, playbooks of FAQs and Agent Skills are written by the customer, and no ready made agents or prebuilt playbooks come with the services.

Sourceredis.io/docs/latest/develop/ai/agent-builderread 2026-11-09

Nothing in Iris starts an agent. Every service answers calls from the customer's own application, which records session events, asks for memories, routes messages and calls tools, while background workers only extract memories and propose playbook entries after the fact. There are no schedules, inbound webhooks or channel connectors that wake an agent. The core database offers Streams and pub/sub that teams use as queues for their own agents, and Data Integration syncs relational changes into Redis within seconds, but neither invokes an agent.

Sourceredis.io/docs/latest/develop/ai/context-engine/conceptsread 2026-11-09

Iris runs models of its own in two places. Agent Memory uses a language model to extract and summarize memories and an embedding model to search them, and Intent Router embeds each route it matches. On Redis Cloud the default is Redis managed keys, where Redis picks both memory models and the customer selects nothing.

With bring my own key, the customer picks a provider, a language model and an embedding model when the service is created, and can later switch the language model or rotate the key, though the provider and embedding model stay fixed. The provider list lives in the console and changes over time, and OpenAI is the one provider shown by name; Intent Router lets each router take a different embedding model. The customer's own agent keeps whatever model it already calls, since Iris runs no inference for it.

Sourceredis.io/docs/latest/operate/iris/agent-memory/model-configurationread 2026-11-09

Python and JavaScript SDKs, REST APIs and the iris command line tool cover every Iris service, for runtime calls and for administration. Each Agent Memory store also serves an MCP endpoint over Streamable HTTP with a bearer key, listing tools that search, create, edit, upsert and delete long term memories, submit documents for extraction and trace memories to their sources. A separate open source Redis MCP server lets assistants such as Claude Desktop and VS Code read, write and query data in any Redis database.

Sourceredis.io/docs/latest/develop/ai/context-engine/agent-memory/mcpread 2026-11-09

No evaluation harness, scored test set or quality gate comes with the services. Lens, still in beta, tags each traced turn as completed, failed, abandoned, refused or timed out and filters sessions by outcome, which shows where an answer went wrong after the fact, and Intent Router returns a similarity score on each match. Versioned playbook entries can be rolled back. Notebooks in the Redis developer guides show RAG evaluation with the third party RAGAS framework, which runs in the customer's own code.

Sourceredis.io/docs/latest/develop/ai/context-engine/lens/conceptsread 2026-11-09

Agents built on Iris reach other systems only through tools behind Context Gateway and the customer's own code, and none of the services drives a browser, a remote desktop or another application's screen. Agent Mesh streams an agent's response events back to a client that reconnects, which keeps output flowing to an interface rather than operating one. The vector search, memory and routing services all work on text and embeddings, and no hosted browser or computer control appears anywhere in the agent stack.

Sourceredis.io/docs/latest/develop/ai/context-engineread 2026-11-09