Video unavailable. Open the source below.
NewThe same typed-decision model supports a complete project-memory loop: classify and filter completed turns, retrieve memories for the next prompt, and evaluate relevant safety rules before tool use.
Jevmem uses Jev to decide what project knowledge from coding-agent turns to retain, whether saved memories should be injected for a new prompt, and whether candidate memories guard tool actions. On a completed turn, it sends scrubbed user text, sometimes a relevant assistant reply, recent context, and keyword-selected existing memories. A broad first tier asks binary Noul questions plus kind and importance judgments; uncertain turns may escalate to a more detailed set of atomic Noul questions. Deterministic policy combines the answers to save, skip, or supersede a memory. Prompt submission separately asks Jev to rank relevant saved memories. The project also uses Jev in guard decisions over candidate rules and tool-call summaries.
Does the user message contain a decision made for this project?
Scrubbed user_message, optional assistant_reply when the turn is a question or reports an attempt, previous_turns (bounded recent context or null), and existing_memories selected by keyword overlap and capped by configuration. Tier 1 asks eleven broad Noul questions when assistant context is included, or ten without it, plus kind Choice, touches_memory_id Choice, and importance Score. Candidate memory choices are dynamic IDs plus none.
Does the user message state a hard rule or limit the project must respect?
Scrubbed user_message, optional assistant_reply when the turn is a question or reports an attempt, previous_turns (bounded recent context or null), and existing_memories selected by keyword overlap and capped by configuration. Tier 1 asks eleven broad Noul questions when assistant context is included, or ten without it, plus kind Choice, touches_memory_id Choice, and importance Score. Candidate memory choices are dynamic IDs plus none.
Does the user message express how the user prefers things to be done?
Scrubbed user_message, optional assistant_reply when the turn is a question or reports an attempt, previous_turns (bounded recent context or null), and existing_memories selected by keyword overlap and capped by configuration. Tier 1 asks eleven broad Noul questions when assistant context is included, or ten without it, plus kind Choice, touches_memory_id Choice, and importance Score. Candidate memory choices are dynamic IDs plus none.
A Stop hook captures completed turns, scrubs secrets, optionally includes relevant assistant text and recent context, and selects a bounded set of existing memories.
Tier 1 evaluates broad content, memory kind, conflict target, and importance questions in one Jev call.
In auto mode, run the larger atomic question set when tier 1 falls within configured borderline conditions; fast mode stays at tier 1 and full mode always runs tier 2.
Combine Jev answers with configured weights and thresholds to save, skip, or supersede a memory; store decision evidence and provenance.
On UserPromptSubmit, rank eligible saved memories against the new prompt and inject relevant results as additional context.
The PreToolUse guard locally prefilters saved rules and asks Jev about the matching scrubbed tool-call summary before enforcing a rule.
The same typed-decision model supports a complete project-memory loop: classify and filter completed turns, retrieve memories for the next prompt, and evaluate relevant safety rules before tool use. The turn path uses a cheap broad pass and escalates only when configured borderline checks indicate uncertainty, with answer caching and provenance logs around calls.
