34 lines
1.2 KiB
Markdown
34 lines
1.2 KiB
Markdown
# Phase 3 Research
|
|
|
|
## Question
|
|
|
|
How should learner memory be added without turning the diagnostic session store
|
|
into durable product truth?
|
|
|
|
## Findings
|
|
|
|
### Keep memory separate from interview sessions
|
|
|
|
Diagnostic sessions contain raw interaction records. Learner memory should be a
|
|
separate derived state built from graded evidence. This preserves the boundary
|
|
between temporary/session context and durable learning claims.
|
|
|
|
### Apply only evidenced updates
|
|
|
|
The memory service should ignore update candidates that do not include evidence.
|
|
This directly enforces the OpenSpec rule that inferred memory requires evidence.
|
|
|
|
### In-memory is still enough
|
|
|
|
The current product does not yet need restart durability. A store interface plus
|
|
tests gives the shape of persistence without adding database complexity early.
|
|
|
|
## Recommendation
|
|
|
|
1. Add `internal/learnermemory` with profile, mastery, misconception,
|
|
intervention, and review schedule records.
|
|
2. Extend workflow runner memory extraction to return evidenced candidates.
|
|
3. Wire diagnostic answer submission to memory ingestion.
|
|
4. Add HTTP read endpoint for learner memory snapshot.
|
|
5. Verify with tests and live smoke.
|