The past conversation and knowledge an AI agent draws from it, such as summaries, facts and user profiles, saved so the agent keeps context from one turn or session to the next. Azure Cosmos DB and Azure Managed Redis are common places to keep it.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Agent memory in context, with comparison tables and the common traps.
Terms in this definition
- AI agent
Software pairing a generative model with tools, enabling it to understand what is asked, converse and act to reach a goal.
- Chat message roles
Labels on chat messages: instructions go under system, the person's input under user, the model's previous answers under assistant, and results returned by a called tool under tool (or function).
- Agent
A specialised form of Microsoft Copilot set up for one particular job, pairing instructions with knowledge and skills. You can create one in Copilot Studio, SharePoint or Agent Builder, and administrators control them from the Microsoft 365 admin center.
- TURN
If a direct link can't be made, RDP Shortpath for Windows 365 relays UDP traffic via Microsoft servers on port 3478 instead.
- NEXT
Restricted to visual calculations, this DAX function reads the value one step further along an axis of the visual matrix; writing OFFSET with 1 gives the same result.
- Cosmos DB
NoSQL database distributed globally, offering writes in multiple regions, automatic indexing and latency below 10 ms.
- Azure Managed Redis
Successor to Azure Cache for Redis: a managed, Redis Enterprise-based in-memory store, with Memory Optimized, Balanced, Compute Optimized and Flash Optimized tiers.
Related terms
- Agent sessions
A Beta feature that keeps the state of a single interaction in Lakebase, typically the chat history or a framework's state like a LangGraph graph. For anything that should persist beyond one conversation, Agent memory is the right place.
- Lakebase
Databricks' built-in Postgres service, run entirely for you and intended for transactional applications. Agents may keep their state there too, but managing your own agent memory in it counts as legacy.
- Self-managed Lakebase
The legacy way to give an agent memory: you create and operate the schema and tables in Lakebase yourself, including connections and access.