redis/agent-memory-server

What it solves

It addresses the scalability and cost issues of using full conversation histories in LLM prompts. Large prompts are expensive, slow, and can lead to degraded accuracy due to bounded context windows, making it difficult for agents to recall information from past sessions without inflating the prompt size.

How it works

Redis Agent Memory employs a two-tier memory model:

• Session Memory: Maintains active conversation state with a configurable time-to-live (TTL).
• Long-term Memory: Stores extracted facts from past sessions as text and vector embeddings for retrieval.

Information is automatically promoted from session memory to long-term memory in the background, allowing the agent to accumulate knowledge without slowing down the live interaction loop.

Who it’s for

Developers building AI agents that require persistent, long-term recall of facts, events, and preferences across multiple user sessions.

Highlights

• Two-tier architecture: Separates immediate session state from long-term factual storage.
• Background extraction: Automatically promotes important information to long-term memory.
• Vector-based retrieval: Uses embeddings to recall relevant facts from previous interactions.
• SDKs and API: Provides Python and TypeScript SDKs and a REST API for integration.