TYPES OF AGENT MEMORY

EXAMPLES Example 1: Short-term memory. A user is chatting with an assistant about planning a trip. The assistant can see the last ten messages in its context window, including that the user wants to visit Lisbon in May. It uses this to answer, "What is the weather like then?" If the conversation grows very long and the early messages are trimmed from the context to make room, the assistant may lose track of the user's budget mentioned at the start. This shows both the usefulness and the limits of short-term memory. Example 2: Long-term memory. A user once said, "Always format my reports as bullet points, not paragraphs." The harness saved this preference in a database tied to the user's account. Three months later, at the start of a brand new session, the harness loads that preference into the context along with a few other relevant preferences. The agent produces a bullet-point report without being asked, regardless of when the preference was originally set. Example 3: Episodic versus semantic. The harness stores an episodic memory: "March 3: user rejected the blue color scheme and asked for green instead." Over time, the user also approves green designs on two more occasions, and once says, "Green is our brand color." The harness then promotes a semantic fact: "The user's brand color is green," marked with its sources: the three episodes and the explicit statement. If the user later says the company rebranded to purple, the system updates the semantic fact and keeps the old episodes as history. Example 4: Cross-session continuity. A developer worked with a coding assistant last week on a project, making decisions such as using a particular database and naming conventions. At the end of that session, the assistant wrote a note summarizing the structure and decisions. Today the developer opens a new session. The harness loads the note, and the assistant begins with, "Last time we chose a relational database and planned to build the login feature next; shall we continue?" The developer does not have to re-explain anything.