Skip to content

Agent memory overview

Every Alfe agent has its own persistent memory. It’s a private, long-term store of what the agent has learned — facts, decisions, preferences, and context from past conversations — that the agent can recall later, even in a completely new session.

Without memory, an agent starts every conversation from a blank slate. With it, the agent can answer questions like “what did we decide about pricing last week?” or “how does this customer prefer to be contacted?” using things it learned in earlier interactions, on a different day, or through a different channel.

What makes it different from a chat transcript

Section titled “What makes it different from a chat transcript”

A chat transcript is a literal, chronological log of messages. Memory is different in two important ways:

  • It’s distilled. Instead of storing every message, memory keeps the durable, reusable facts — the things worth remembering — and leaves the small talk behind.
  • It’s semantic. You retrieve from memory by meaning, not by keyword. Asking “how does Acme like to be billed?” can surface a memory that was originally phrased “Acme wants a single invoice at month end,” even though the words don’t match.
  • Per-agent and private. Each agent has its own memory. One agent’s memories are never visible to another agent. Memory belongs to the agent, not to a single conversation.
  • Persistent across sessions and channels. Something an agent learns while chatting on the web is available later over Slack, SMS, voice, or any other channel the agent is connected to.
  • Automatic. For the most part you don’t manage memory by hand — the agent builds and recalls it as it works. See How memory works.
  • Organized. Memories are grouped under a simple topic → subtopic → tag hierarchy so related knowledge stays together and can be browsed or filtered.
  • Programmable. When you’re building on the agent, you can read and write memory directly through the Agent API.