The Architecture of Synthesis

In brief

What this examines

A useful AI answer can still lose the context that made it useful. The Architecture of Synthesis explores a practical exocortex: an external layer of memory and reasoning that carries decisions, evidence, and limits forward between sessions. Its proposed architecture combines durable records, checks at the point of action, observed outcomes, and independent review so that persistent context remains open to correction.

Why it matters

Persistent AI environments can carry old assumptions into repeated action. Their reliability depends on checking whether sources, decisions, and approvals still apply, enforcing the limits of granted authority, and preserving evidence that lets people question the result.

Key ideas

• Decision records should preserve what was decided, why, on whose authority, from which evidence, and when the source was last checked.
• An enforceable action check must sit outside the acting model's discretion, cover the relevant action path, and resist bypass or disabling by that model.
• A successful tool call is not sufficient evidence of completion. Inspect the outcome and preserve uncertainty before retrying an action.
• Make requirements inspectable while retaining substantive review: a citation can exist without supporting a claim, and a complete document can still explain its subject poorly.
• Review memory against current sources, observed outcomes, independent checks, and human judgment. Corrections need authority and should preserve unresolved disagreement.

Carry context forward

The article uses exocortex to describe an external layer of memory and reasoning that stays attached to the work over time. Ordinary components such as durable records, event logs, access controls, workflow software, and language models can support that environment.

Useful records preserve the reasoning and evidence behind decisions, along with their authority and last source check. A new request can draw on that context to form a proposal, which still needs a separate authority check before action.

Check authority and verify the result

Written rules guide a model; enforcement requires a control where the action occurs. A hook qualifies as an enforcement mechanism only when the relevant action must pass through it and the acting model cannot bypass or disable it. Missing evidence or an unavailable check should stop a consequential action.

The article illustrates this with a hypothetical customer report whose export gate checks the recipient, data classification, and required approval. Approval must cover the actual report and recipient. After execution, the system should inspect the outcome and record what it observed, including uncertainty that could make a retry unsafe.

Make judgment inspectable

Experienced people recognize requirements that may never have been written down. Making those requirements explicit helps software check whether claims follow current sources, an edit stays within scope, another writer's work is preserved, and the final artifact is usable.

Mechanical checks and substantive review serve different purposes. Evidence must support the claim beside it, and a document needs to explain its subject well. The people responsible for the work retain authority over the conditions for success.

Keep the review loop open

Persistent memory needs evidence from outside its own account of events. Current sources, observed outcomes, independent checks, and human review can reveal stale policies, conflicting instructions, or approvals that cover an earlier version.

Meaningful review presents the proposed change, supporting evidence, applicable limit, and remaining uncertainty. An accountable owner decides which sources deserve weight and who may change a rule. Authorized corrections update the record while preserving unresolved disagreement.