Skip to main content
This document introduces how AgentScope handles context and memory in agent workflows.
For implementation details and APIs, see Context and Memory.

​
Why It Matters

Without memory, an agent treats every turn as a new conversation. With memory, it can:
  • keep conversation continuity,
  • remember user preferences and task state,
  • retrieve useful history when needed.
In AgentScope, this is built around Msg, memory backends, and prompt construction.

​
Three Layers in AgentScope

​
Context (inference-time input)

Context is the final input sent to the model for one inference call. In AgentScope terms, it is typically assembled from:
  • system instructions,
  • current user Msg,
  • selected short-term history,
  • retrieved long-term memory,
  • tool results.
Context is fast and direct, but limited by token budget.

​
Short-Term Memory (session state)

Short-term memory tracks the current session and usually stores Msg objects. In AgentScope, this is provided by MemoryBase implementations (for example: InMemoryMemory, RedisMemory, AsyncSQLAlchemyMemory). Common usage:
  • recent conversation turns,
  • temporary task progress,
  • marked messages (such as hint, summary, tool_result).

​
Long-Term Memory (cross-session knowledge)

Long-term memory stores information that should survive session boundaries. In AgentScope, this is abstracted by LongTermMemoryBase implementations. Typical content includes:
  • stable user preferences,
  • important facts from previous interactions,
  • retrievable semantic memories.

​
How They Work Together

A typical agent turn looks like this:
  1. Load recent session memory (short-term).
  2. Retrieve relevant long-term memory (if needed).
  3. Build model context with current user input and retrieved signals.
  4. Run model inference.
  5. Write new messages back to short-term memory, and optionally persist key facts to long-term memory.
This loop is often managed inside the agent’s reply lifecycle (for example, in ReActAgent workflows).

​
Context Management vs Memory Management

In practice, the boundary is soft:
  • Memory management focuses on storing, retrieving, marking, and updating information.
  • Context management focuses on selecting and assembling the right subset of that information into the model input.
So memory provides the source material, and context decides what enters the current inference window.

​
Summary

LayerWhat it isAgentScope mapping
ContextInput for one inferencePrompt assembled from Msg + retrieved memory + tool outputs
Short-Term MemorySession-level working stateMemoryBase backends storing and filtering Msg
Long-Term MemoryPersistent cross-session knowledgeLongTermMemoryBase retrieval and storage
Good agent behavior depends less on storing everything, and more on selecting the right memory at the right time.