Redis Buys Decodable Because AI Agent Memory Is a Mess - September 5, 2025

Redis is buying Decodable because anyone who's tried building AI agents knows the painful truth: they forget everything constantly. LLMs are stateless. Context windows are tiny. Getting fresh data to agents in real-time is a nightmare of custom pipelines that break constantly.

Decodable solves the "get data from everywhere into Redis fast" problem that makes AI agents usable in production. Instead of spending weeks building custom data pipelines that sync customer data, transaction history, and real-time events into your agent's memory, you just configure Decodable streams. It handles the painful stuff automatically.

CEO Rowan Trollope gets it: "The challenge isn't proving what language models can do; it's giving them the context and memory to act with relevance and reliability." Translation: your customer support agent is useless if it can't remember previous conversations or access current account status.

The real problem is that AI agents need memory that works like human memory - instant access to recent interactions, ability to retrieve relevant context from history, and fresh data about what's happening right now. Most developers cobble this together with a mess of databases, APIs, and custom code that fails in production.

LangCache Actually Saves Money (Finally)

Redis also launched LangCache, which is semantic caching that actually works. Key stats that matter:

• Claims big reduction in LLM API costs - this could add up fast when you're burning hundreds on OpenAI calls
• Much faster responses when cache hits - users hate waiting 3 seconds for ChatGPT
• Semantic matching - "What's the weather?" and "How's the weather today?" hit the same cache

Normal caching sucks for LLM queries because nobody asks the exact same question twice. Semantic caching understands that "Show me my recent orders" and "What did I buy lately?" mean the same thing. This is the kind of obvious feature that should have existed years ago.

Framework Integration (Because Nobody Wants Vendor Lock-in)

Redis finally realized developers don't want to learn proprietary APIs. New integrations include:

• AutoGen - Microsoft's multi-agent framework with Redis memory
• LangGraph - persistent memory for agent workflows
• Cognee - memory management with summarization

Smart move. Developers already know these frameworks. Redis provides the fast memory layer underneath without forcing you to rewrite everything. The memory problem is hard enough without learning new APIs.

The Decodable acquisition is Redis betting that developers are tired of building data pipelines from scratch every time they want AI agents with fresh context. And they're right. Most AI agent projects die in the "getting real-time data into the agent" phase.

The Data Pipeline Hell Problem

Here's what building AI agents looks like right now: You want a customer support agent that knows about recent orders, support tickets, and account changes. So you spend weeks building Kafka consumers, writing ETL jobs, and debugging why your data is always 30 minutes behind when users complain.

Decodable turns that into configuration files. Instead of custom Apache Flink jobs that break when someone touches them, you declaratively define "stream customer data from Postgres, order data from the orders API, and support tickets from Zendesk into Redis." It handles the plumbing.

The combination solves the "stale context" problem that makes AI agents useless in production. Your agent needs to know the customer just placed an order 5 minutes ago, not what was in the batch job from last night.

Performance Numbers That Actually Matter

Redis also dropped some performance improvements that matter for real apps:

• Redis 8.2 is 35% faster with 37% smaller memory footprint - real money savings if you're paying cloud bills
• Int8 quantized embeddings cut vector search costs by 75% and run 30% faster - finally makes vector search affordable at scale
• Hybrid search improvements for better text + vector ranking

That quantization improvement is huge. Vector search was expensive enough that most companies couldn't afford to do it properly. 75% cost reduction makes it feasible for normal apps, not just AI unicorns with unlimited budgets.

Competition Reality Check

Redis is fighting on multiple fronts. Pinecone owns vector search. Weaviate and Chroma have strong open-source communities. But Redis has one major advantage: you're probably already using it.

If you're running Redis for caching and sessions, adding AI agent memory doesn't require new infrastructure. You don't need to learn another database, manage another service, or justify another vendor to procurement.

The cloud providers are building Redis competitors (AWS MemoryDB, Google Memorystore), but they're still catching up on AI features. Redis has a head start on semantic caching, agent framework integrations, and AI-specific optimizations.

The timing makes sense. AutoGen is getting real enterprise adoption. LangGraph powers a lot of production agent deployments. Companies are moving from "let's try building an AI agent" to "we need this agent to actually work in production." That's when memory and data freshness become critical problems.