The memory stack
Seven layers, numbered by the timing budget each consumes relative to a conversation turn.
L0–L4 is per-agent, the recognition substrate each adapter holds locally. L5–L6 is shared across every adapter on the machine, exposed as an MCP server. Recognition fires from L0 in milliseconds, hydration runs in parallel from L1, and archive descent only happens when the conversation demands it.
Click any layer for the full description.
The numbering encodes timing budget, not priority. A turn that resolves at L0 entirely is the fluent case. Descending to L3 or L4 is deliberate and rare. The federation layers (L5, L6) are what let agent continuity survive vendor account and machine boundaries, see the cross-account and cross-machine field tests for measured outcomes.