Skip to content

ADR-0206: Default deliberate memory search to the trusted session

  • Status: Implementation decision under the owner's recall-release repair request
  • Date: 2026-10-15
  • Related: ADR-0178, ADR-0198, ADR-0199, ADR-0205
  • Detailed design: docs/plan/general-memory-recall.md

Problem

The retained failed confirmation on 74b14815 contains a mandatory incident failure: a formed fact reached deliberate recall, but the reader requested clarification. Its search used a different scope from its session. That difference changes portability attribution and source recovery. It does not by itself prove why the reader declined to answer. The production tool currently requires the model to supply a scope even though the session already establishes one.

The same attempt also exposed diagnostic loss: a structured provider failure can be overwritten by the evaluation wrapper's missing-final-usage exception. Its uncertain charge remains reserved, but the original safe error code is lost.

Decision

The model-facing memory.search@2.1.0 makes scope optional. Omitting it resolves the authenticated session's project through the existing persistence port. The description directs ordinary searches to omit scope; an explicit scope retains its existing meaning for intentional cross-project lookup. Preparation and tool execution resolve the same default, so a prepared query cannot silently differ from its later search. Failure to resolve the session fails the operation; it does not invent a default project. The legacy direct tool keeps its required scope and original version. Bootstrap retains its 1.1.0 registration for existing pinned sessions and advertises only 2.1.0 to new sessions. Recall preparation resolves that same pinned version, preserving the legacy hard filters and reserving an execution allowance for valid old calls. No principal, sensitivity, portability, lifecycle, confidence, source-recovery, or erasure requirement changes.

The evaluation budget wrapper passes a failed terminal event through to the normal stream collector and journals only its safe classification, provider code, parameter and HTTP status. It retains the full uncertain reservation and refuses further paid calls. It adds no retry and cannot convert a failed observation into successful evidence. Provider messages and partial content stay out of the journal.

Verification and release boundary

Regressions cover default scope on multiple sessions, deliberate preparation and execution parity, retained explicit cross-project lookup, compatibility resolution and hard filtering for pinned legacy sessions, and provider failure identity with uncertain accounting. These mechanism tests do not establish reader quality or close the failed incident gate.

The owner requested completion of the remaining gates before merging PR #163. The original failed attempts and their identities remain immutable. Aggregate reported and held costs are USD 55.913546; a fresh complete confirmation may use at most USD 240 after rechecking the ledger, within the existing USD 300 cap. Changed evaluated inputs require new matching measurements, final CodeRabbit review, exact-head hosted verification and signing before that paid run. All quality, cost, latency, independent human review and publication requirements remain unchanged. No optional channel or unattended dreaming is activated here.