In-band memory writing

In-band memory writing is the process by which an AI agent writes to its own persistent memory store during task execution, as opposed to offline or out-of-band memory consolidation processes. It is contrasted with Dreaming|dreaming, an offline approach to memory refinement.

How it works and its improvement over model generations

Lance Martin (Anthropic) observes that Claude Code|Claude has improved substantially at in-band memory writing across successive model generations. (3:04:40) A key driver of this improvement is model capacity: higher-capacity models develop a better sense of what level of abstraction to commit to memory — saving generalizable facts and patterns rather than task-specific details that will not transfer across sessions. (3:27:40)

Schema and structure

Martin advocates letting the model structure and maintain its own memory, rather than supplying a prescribed memory schema. Imposing a fixed schema degrades performance because it is difficult for humans to anticipate ahead of time what memory organization will be useful; models manage their own memory structure better when unconstrained. (6:29:30)

Points of disagreement

Reliability and error persistence

Despite its improvements, in-band memory writing carries a structural vulnerability: because writing happens during task execution, errors can be committed to memory and persist indefinitely, gradually degrading agent performance. Martin notes that "those mistakes get stuck in memory unless you have an offline process to kind of correct them." (4:26:40)

The proposed remedy is a complementary dreaming process — an offline, out-of-band routine that reviews prior session traces and corrects the memory store after the fact. (3:39:50) Martin thus positions dreaming not as a replacement for in-band memory writing but as a necessary correction layer on top of it; in-band writing is fast and continuous, while dreaming handles consolidation and error repair that the in-band process cannot reliably perform on its own.