NewEkko Studio
Optional Jev checks filter and rerank memory cards for an agent workspace.
Added to Jevfast
NewHow it uses Jev
When the memory relevance filter is enabled, Ekko asks one Choice question for each unprotected candidate memory card. Each question compares that card's actual information with the current request and offers the source-defined choices relevant and irrelevant. Ekko removes a card only for a confident irrelevant answer; uncertain judgments leave it available. A separate optional pass scores candidate usefulness for reranking before the existing token selection.
What Jev decides
Ekko memory-card relevance (illustrative request)Example answers · not a recorded Jev response · Source ↗
Question 1 · filter 0
YOUR APP
INSTRUCTION
Does the actual information in this card help answer the current request, including by correcting a false premise?
STATE
Source-defined JSON string with request and bounded cards containing kind, title, content and value. This representative chart shows filter_0 for an unprotected first candidate; optional kind-routing questions share the same call. No captured request or answer is claimed.
JEV · CHOICE
- relevant
- irrelevant
Ekko memory usefulness ranking (illustrative request)Example answers · not a recorded Jev response · Source ↗
Question 1 · card 0
YOUR APP
INSTRUCTION
How useful is this card to the current request?
STATE
The request contains the query and bounded candidate cards represented by title, content, and value. One score question is generated per candidate card. This representative card_0 refers to the first of at least two candidates; no runtime values are claimed.
JEV · SCORE
0
App workflow
- Prepare candidate memories
Use eligible candidate cards and the current request; protected cards are excluded from relevance filtering.
- Filter by request relevance
Ask Jev about each unprotected card and remove only cards marked irrelevant at or above the configured filter-confidence threshold.
- Rerank candidates
When enabled and enough candidates remain, score the bounded candidate set and order it by score, preserving original order for ties.
- Use memory context
Continue the existing memory recall and token-selection flow; scores do not become stored card fields.
Why it is interesting
The filter tests whether a memory helps with the actual request, rather than treating shared words or a broad preference category as enough. Its policy also preserves protected memories and lets uncertain judgments abstain, while reranking keeps provider scores internal to ordering.