Agent Memory Isolation

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

Obtain two independent initial auth-change reviews. Teaching A/B share revision/requirements but not draft notes; approved completed reports use a separate shared entrypoint. A private folder name alone prevents no reads.

Mechanism

Define common inputs, private working state and exchange timing. Use supported isolation/access contracts with complete facts and no preferred verdict. Collect completed initial reports before disagreement discussion. Inspect access/state and unintended cache/file sharing.

Bad example

Read all peers' drafts before analysis, treating matching conclusions as independent confirmation. Call private directory naming access enforcement.

Good example

A/B independently review one revision/requirements without reading private drafts. Verify actual host limitations or explicit cooperation rules. After completion exchange permitted shared/reports content, retaining evidence/disagreement without anchoring on the first opinion. Inspect sequence/content rather than claiming forced isolation from names.

Why the change matters

Separate state reduces direct verdict reuse while controlled exchange supports evidence integration. Inputs/access records make independence inspectable rather than a folder-label claim.

Observable expectation

Initial report timestamps precede cross-reading; shared material excludes private notes. Inspect actual access/protocol and disclose premature draft reading instead of claiming isolation.

Fresh context establishes neither unbiased nor correct judgment; no real private memory is handled.

Limits

Some hosts provide cooperative rather than enforced boundaries. Candidate scoring/project scope are related instances; this pass adds direct frozen independent-review-context evidence with a separate record.