Daniel
Sets intent, resolves ambiguity, and approves consequential work.
DECIDESAGENT MEMORY & RELIABLE ACTION Architecture overview · August 2026
A visual guide to why we are pairing active agent memory with deterministic business controls—and how a year of operating experience moved us here.
The architecture
The upper half shows where flexibility belongs. The lower half shows how operating experience produced the design.
How it operates
The agent can reason flexibly, but work, authority, and results move through registered components. Operational state lives in structured records—not a pile of .md files that can drift apart.
Sets intent, resolves ambiguity, and approves consequential work.
DECIDESManages attention: what changed, what needs a decision, and what is blocked.
PRESENTSOwns project registry, scope, routing, ownership, lifecycle, and approval state.
AUTHORIZESInterprets the task and performs only the work allowed by the registered route.
PROPOSES + ACTSExposes typed reads and narrow approved writes. It is a boundary, not a second brain.
CONSTRAINSStores structured projects, tasks, work logs, source observations, and lifecycle state.
RECORDSSupabase result
←MCP response
←Control Center lifecycle
←Desk status
←Daniel sees the outcome
The agent’s understanding does not grant permission. Control Center scope and approval govern what may happen.
A write or external action is not complete until the result is read back and reflected in lifecycle state.
.md FILESUseful for plans, explanations, and exports—but not the operational database, routing registry, or current status.
The progression
This is the short version. The detailed page preserves the reasoning, limits, and supporting evidence.
Section guide
Each card gives the point in plain language and leads to the exact section containing the full reasoning.
Use flexible navigation to decide what evidence matters. Use deterministic controls to decide what actions are allowed.
Read the position → 02THE SYSTEMKnowledge, governance, execution, and presentation cooperate without silently inheriting one another's authority.
See the system → 03HOW WE GOT HEREAttempt 5.5, Control Center, GLOBAL, MCP, the Desk, and email intake each revealed the next missing boundary.
Follow the chain → 04THE CUSP CLAIMThe components are established or emerging. Their disciplined use across live business operations is still early.
Test the claim → 05THE EVIDENCEMemGPT, MemOS, and NapMem inform the memory path. Current agent guidance informs governance and oversight.
Review the sources → 06THE CRITIQUEPrior art, counterexamples, simpler architectures, and failure modes are more useful than agreement.
Send feedback ↗FULL RECORD Argument · decision chain · evidence
Read what we adopted, what we deliberately rejected, what remains unresolved, and what evidence would cause us to change direction.