The memory layer for AI agents

Context Streams
for AI agents.

The context layer for AI agents. Like Vercel KV but for agent memory. Publish interactions, query by key or vector, get materialized context. No Kafka required.

Learn More Get in Touch
# Publish a message with embedding curl -X POST https://streams.protobuf.ai/streams/support/messages \ -H "Content-Type: application/json" \ -d '{ "key": "customer-123", "data": "eyJpc3N1ZSI6ICJjYW5ub3QgbG9naW4ifQ==", "embedding": [0.1, 0.2, 0.3, ...] }' # Response {"message_id": "msg_8x7k...", "sequence": 42}

Architecture

Three layers. Real-time context. No infrastructure nightmare.

┌──────────────┐ │ AI Agent │ └──────┬───────┘ │ ▼ ┌──────────────────────────────────────────────────────────────┐ │ Context Streams API │ │ POST /messages GET /context/{key} POST /search │ └──────────────────────────────────────────────────────────────┘ │ ▼ ┌──────────────────┐ │ PostgreSQL │ │ + pgvector │ └──────────────────┘

Built different

Not another Kafka. Purpose-built for AI agent context.

⚡ Instant Context

GET /context/{key} returns materialized context in milliseconds. No query building. No joins.

🔍 Vector Search

Semantic similarity search across your context. Find relevant past interactions, not just exact matches.

📊 Typed Context

Schema-defined context structures. No untyped JSON blobs. Your agent memory has contracts.

🔒 Durable by Default

PostgreSQL handles ordering and durability. Your messages won't disappear.

🚀 No Infrastructure

No Kafka clusters. No Zookeeper. No schema registry configs. Just context that works.

🔌 Just HTTP

REST API. Works from any language. No client libraries required. curl it.

Why agents need context streams

MCP gives agents tools. But where does memory live?

The Problem

AI agents can use tools via MCP. They can call APIs. But they have no persistent, typed, searchable memory. Every conversation starts from zero.

The Hack

Teams stuff context into system prompts. They serialize JSON to vector DBs. They build bespoke context pipelines. None of it is typed. None of it is standard.

The Solution

Context Streams: schema-defined memory with vector search. Publish interactions with embeddings. Query by key or similarity. Get materialized context.

Give your agents memory

Typed context. Vector search. No infrastructure.

Get in Touch