Agent memory

Agent memory is how an agent keeps information beyond a single request. A language model doesn't remember anything on its own: it builds each answer only from what's in its context window. So agent memory is data the system stores and adds to the context when it's needed.

Types of memory:

• Short-term: the current conversation history in the context window
• Long-term: facts and preferences stored between sessions
• External: documents and a knowledge base the agent pulls from via search (RAG)

Too much memory backfires: it takes up context and confuses the model. That's why an agent only gets what's relevant to the task.

In Pachca's breakdown of its own agent architecture, short-term memory is the last 20 messages in the thread plus the parent message that started it. Personal settings, like “always reply to me in English,” are stored in the user's Pachca profile and applied to every request. And with a link to a thread, an agent can read a past discussion in full.