Substrate memory vs. Anthropic managed memory

Reported results from real deployments are substantial: Netflix carries context across sessions, Rakuten’s long-running task agents report 97% fewer first-pass errors from using memory to avoid repeating past mistakes, and Wisedocs built a document-verification pipeline on cross-session memory. These are legitimate outcomes for one agent’s own continuity, not a shallow feature. Primary source: Anthropic’s Managed Agents memory announcement.

Managed Agents sessions are isolated by design: each session’s memory store is its own, and nothing shares between agents natively. Pooling experience across multiple agents requires adding a separate vector database outside the shipped memory mechanism. The feature also runs inside Anthropic’s own hosted environment; it does not extend to agents built on other model vendors or other runtimes. Read together, the shape is memory scoped to one agent’s own container, which is what isolation by design means.

A model lab’s product is the model. Shipping a persistent, cross-session, cross-agent, cross-vendor substrate would mean building and operating infrastructure whose center of gravity sits outside the model itself, a different kind of product with a different set of guarantees. That is not a matter of engineering effort so much as a constraint on what a model lab is organized to build.

The same logic separates a workflow runtime’s intermediate state from a chat’s context window: keeping a session’s memory scoped to that session and that container is what keeps the product legible as the model plus its own history, rather than a second, separately governed product line.

Substrate memory does not change how Managed Agents memory behaves; a session reading and writing its own /mnt/memory/files continues to do exactly that, unmodified. It does not replace the mechanism Anthropic ships; the two operate at different layers. And it is not a parallel duplicate of it either: it adds a layer that a session-scoped, single-agent memory store does not attempt to cover, a graph that persists across sessions and across the agents and clients that write to it, kept outside any one vendor’s runtime.

Dynamic Workflows illustrate the same shape at a shorter timescale. A workflow’s subagents can use well over a million tokens of intermediate reasoning that never enters the orchestrator’s context, and only the converged synthesis returns to the chat. The workflow runtime does not remember having converged once the run ends; persisting that result across sessions requires a store outside the runtime that produced it. Substrate memory names that layer, and treats each new session-scoped primitive a model lab ships as another source that can write to the graph, not a competing store to reconcile against it.

Managed Agents memory addresses part of the third row: one agent’s own continuity, inside one vendor’s hosted runtime. A graph that persists across sessions, across agents, and across vendors addresses the same shape at every row, a write target that sits outside whichever runtime produced the reasoning in the first place.

The two are not mutually exclusive. Managed Agents memory suits an agent’s continuity across its own sessions inside a hosted Claude environment: a shipped, working answer to whether an agent already tried something. Substrate memory suits reasoning that needs to persist across sessions, across the several agents and clients that touch a shared piece of work, and across whichever model produces a given turn. A team can run both: one agent keeping its own /mnt/memory/ continuity inside Managed Agents, while every agent involved also reads and writes a shared graph that outlives any single session or vendor.