Typed Memory Taxonomy

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

Use case

Reuse a user’s test preference, task progress and a deployment endpoint. Teaching inputs say ‘For this project, run unit tests first’; the task’s checks remain unfinished; the endpoint comes from current deploy/config.json. One permanent memory paragraph can confuse preference, state and changing facts.

Mechanism

Classify records by purpose: explicit feedback with project scope, task state with plan/revision, and a reference to changing endpoint configuration with observation time. Verify by type: check superseding preferences, state artifacts/revisions and current configuration. Persist only authorized necessary content; silence is not confirmation.

Bad example

Save the preference, unfinished progress and current endpoint as permanent facts for every project. Trust them forever without rereading configuration.

Good example

Classify the supplied inputs in authorized records: feedback—unit tests first for this project, from the explicit user request; task-state—checks unfinished, bound to plan/revision; reference—deploy/config.json determines the endpoint, with location/time. Verify by type on resume or release preparation. Reread the endpoint; honor newer user requirements; do not turn unfinished work into completion. Store no credentials and do not treat silence as confirmation.

Why the change matters

The records have different validity and uses: preference affects order, state affects continuation and configuration supplies current facts. Classification enables targeted verification rather than treating a stale endpoint as durable preference or promoting one task state into a global rule.

Observable expectation

After changing the teaching endpoint in deploy/config.json, preparation should use the new value. A newer ‘inspect compilation first’ request supersedes the relevant feedback. Unfinished checks cannot become complete through an older report.

Inspect each record’s type, scope, source and verification rule. Cross-project preference reuse or skipped configuration reads fail the check.

Limits

These types are a teaching organization, not a universal memory schema. Record access and persistence remain authorization-bound. Labels confer neither trust nor release permission. The source remains unconfirmed, and current evidence outranks stale notes.