The Brain · 04.01
One Substrate, Many Planes
All Brain memory lives in one table, discriminated by plane, scope, and type — why a single substrate beats a dozen disconnected memory stores.
One table, not a dozen half-integrated stores
Every durable memory in Agentis — a rule an operator stated in chat, a lesson distilled from a failed run, a fact imported from a harness file, a skill's confidence score — lives in the same underlying table, memory_episodes. A typed facade (MemoryStore) keeps the API ergonomic, but underneath it there is exactly one index. That matters more than it sounds: recall, decay, analytics, and maintenance all read the same rows, so a fix to how recall ranks results improves every kind of memory at once instead of needing to be reimplemented per-feature.
Discriminated by plane, scope, and type
| Dimension | Values |
|---|---|
| Plane | workspace_memory (dispatch-injected durable context) and skill_library (searched or materialized skill knowledge) |
| Scope | Workspace-global, or bound to a specific agent, workflow, or app |
| Type | Observations, success patterns, distilled lessons, decisions, failure chronicles, and other outcome-tracked record kinds |
Scope is what makes a workspace-wide convention and an individual agent's private preference coexist safely: a candidate memory carries a scope hint at write time (see The Formation Pipeline), and recall only ever surfaces what a given scope is actually allowed to see.
Continue
The idle watchdog, turn deadline, budgeted context window, and tool cap that keep a chat turn from running away.
Writing memory is reconciliation, not append: deterministic extraction, a two-phase Formation Judge (ADD/UPDATE/NOOP), PACER routing, and Feynman repair.