Working memory
Working memory is the agent's view of the current conversation: the last N turns, the immediate context. It lives in the prompt window and turns over as the conversation moves.
Why it matters
Working memory is what makes a chat feel coherent. If the agent forgets what you said three turns ago, the conversation breaks down. The size of the working memory window depends on the underlying model's context window plus the system's allocation logic; in 2026 a typical working memory carries 10-30 turns.
The interesting design problem is what to do at the working-memory boundary: what gets promoted to episodic memory versus what gets discarded as filler.
How Ajeris handles this
Working memory in Ajeris carries the most recent 10 turns of the active conversation, plus the relevant slice of cross-channel turns from the past 30 minutes. As turns age out of the window, important content is promoted to episodic memory automatically. The agent's prompt is rebuilt on every turn from working + episodic + semantic + core; the user does not see the layer boundaries.
