Agent topics

... or component to consider

Frameworks here refers to such frameworks as Google ADK, LangChain/Graph, Pydantic AI or Smollagents.

• Memory
  • Conversational
  • Long term
• Context
• Authorisations
• Goal/mission

Remember all memory seen by the LLM, must fit in the context window and will contribute to the amount of input tokens.

Conversation memory

For contextual message, like saying "go ahead, do just that". The LLM needs a history of what went before. This kind of memory is usually support diretly by framworks.

Conversational memory, is bound to a single session. What delimits a session, is up the agent implementation. It could be, "until user clicks new conversation" or after some idle time.

For agent deployment with multiple processor conversation memory must be in a persistent shared store.

Long term memory

Long term memory refers to memory of past conversations.

It can be benefical to have an agent remember past conversation. Such memory should be in a summarised form.

Long term memory must be in persistent store.