Jcode memory recall
Select relevant coding-agent memories with independent Jev judgments instead of embeddings.
Added to Jevfast
How it uses Jev
jcode sends the user's memory-recall query and batches of active memories to Jev. Each candidate contains only its content, category, and tags, and gets an independent Noul judgment about whether it directly helps answer or act on the query. jcode keeps candidates whose returned probability meets the configured relevance threshold, then ranks them and applies the requested result limit before using the selected memories for recall.
What Jev decides
Judge whether a memory helps answer the queryExample answers · not a recorded Jev response · Source ↗
Question 1 · candidate 0
YOUR APP
INSTRUCTION
Judge this candidate independently: is it directly relevant and useful for answering the query? Broad overlap is insufficient; prefer false when relevance is uncertain or incidental.
STATE
For each bounded batch of active memories, state contains the user's query and a candidates map. Each candidate has content, category, and tags; candidate_0 is a representative positional reference, and its memory content and metadata vary with the stored entry. The example describes request shape, not a captured Jev answer.
JEV · NOUL
YesNo
App workflow
- Collect active memories
Load the requested project and/or global memory scopes and collect active entries; storage errors propagate instead of becoming an empty result.
- Build bounded requests
Sort entries into a canonical order, then batch up to 24 candidates. Include only each candidate's content, category, and tags; skip an individually oversized entry rather than truncating it.
- Ask Jev for relevance
For each candidate, send a Noul question that asks whether it directly helps with the query. Process batches sequentially under a total deadline and validate the complete answer set.
- Filter and return memories
Keep scores at or above the configured threshold, sort by relevance while preserving canonical tie order, and truncate to the requested result limit.
Why it is interesting
The recall path replaces embedding search and generative recall with typed relevance judgments over every active memory. It keeps provenance and other internal fields local, uses stable ordering for ties, and fails the selection as a whole when a request or response is invalid.