How-to · Search
How to answer project questions with verified memory
Use the search layer to retrieve the right tasks and decisions before an agent writes a single line of code.
Start with the question, not the files
vem is optimized for question-led retrieval. Query for the intent first, then link to the exact tasks and decisions that provide the context.
This avoids the common trap of scanning docs without knowing which artifacts actually matter.
Strong queries are short, explicit, and tied to a project outcome.
- Who approved the auth model?
- What tasks shipped the org switcher?
- Which ADR governs retention?
Prefer verified snapshots
Verified updates provide the canonical record, so they should rank above pending changes in search.
That keeps agents anchored in decisions that have passed review.
When in doubt, show both but clearly label which memory is verified.
Close the loop with updates
Once a question is answered and work is done, update the task or decision so the next retrieval is better.
This continuous improvement loop is what makes memory compound over time.
A tight feedback loop keeps the search layer crisp and reduces rework.
- Mark tasks done with evidence
- Summarize decisions after approval
- Update context packs routinely
Examples of high-signal questions
Which decisions led to the current authentication model?
What tasks were completed to support organization switching in the web app?