Memory substrate

concept · updated Jul 25, 2026

person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation

A memory substrate is the underlying storage layer (e.g., a database or file system) that an AI agent uses to persist and retrieve information across a long-horizon task. The key design question is how much structure the developer imposes on that storage versus how much the model manages itself.

General-purpose substrates outperform rigid schemas

Lance Martin (Anthropic) argues that flexible, general-purpose substrates — plain databases or file systems — consistently outperform tightly structured ones because they allow the model to organize memory in whatever way suits the task at hand:

"what we've seen is that very general substrates for memory, be it just your database or file system, are good because the model can manage them freely" — Lance Martin, 6:08:40

The corollary is that developer-imposed schemas hurt performance. Martin specifically criticizes approaches that force a model into a predefined memory structure:

"a very very kind of like prescriptive memory schema they are trying to pigeonhole the model into. That's when you see performance drop" — Lance Martin, 6:16:20

The underlying principle is that models are often better judges of how to structure their own working memory than developers can anticipate ahead of time; pre-specifying a schema removes that flexibility and introduces a mismatch between how the model "thinks" and how it is allowed to store information.

Design implication

The practical takeaway for agent engineers is to resist the temptation to define a detailed memory ontology upfront. Instead, expose a general substrate (a key-value store, a file system, a schemaless document store) and let the model determine what to write, how to label it, and how to retrieve it — intervening with structure only where there is strong empirical evidence it helps rather than hurts.