Memory Management

Memory management governs when agents store new memories, how they retrieve relevant ones, and when they evict or summarize old information to stay within operational limits. The practical decision you face at every context boundary is not whether to save information but how: summarization compresses a long conversation into a durable paragraph the agent can reference later, truncation simply drops the oldest turns and accepts the loss, and explicit forgetting removes entries that are outdated or contradictory rather than just old. Choosing wrong here produces the most common agent failure pattern: an agent that truncates mid-task loses the constraints set at the start of the session and spends three steps rediscovering them, while an agent that never forgets fills its context with stale state that actively misleads its next decision.