Knowledge Planes: Why Agents Should Share Understanding, Not Keep Memories

Much of the LLM-agent field has converged on a memory architecture: a private or product-owned store of extracted facts and episode summaries, written automatically and retrieved at inference time. In the systems surveyed here, that store is rarely reconciled against the curated knowledge beside it: the agent interprets the world through one substrate while its experience accumulates in another, and the two carry different warrants and usually meet only when retrieval places them in the same context window. Computational accounts of biological memory rely on a consolidation pathway between fast episodes and slower structure; the surveyed agent stack lacks an equivalent governed, content-level pathway for declarative claims. We propose a multi-plane knowledge substrate in which memory is demoted to an unconsolidated ingestion buffer, and everything durable is knowledge held at some plane (task, user, team, organization, or global), each with its own access policy, editorial policy, lifetime, and purpose labels. Consolidation becomes promotion-with-review across planes, enforced by a per-plane write gate. We develop four governance constraints that follow from that design: access control is not consent, so purpose must be enforced rather than inferred from identity; a shared plane is a shared poisoning surface; disagreement can be evidence rather than a synchronization error; and a user plane must remain under the user's logical control even when a vendor hosts it. The payoff is agents that can share reviewed understanding without turning every recollection into common ground.