NewJev-Mem
Uses fast decisions to organize memory and guide retrieval for long-running agents.
Added to Jevfast
NewHow it uses Jev
Jev-Mem puts typed Jev decisions around an agent memory graph. On a new observation it batches admission and memory-type Noul questions; for a user query it routes retrieval and tests whether gathered evidence is sufficient before a separate language model writes an answer.
What Jev decides
Observation admissionExample answers · not a recorded Jev response · Source ↗
Question 1 · should store
YOUR APP
INSTRUCTION
Does this observation contain a detail worth retaining for future recall?
STATE
observation text, exact_duplicate flag and recent_memories list from the memory policy.
JEV · NOUL
YesNo
Question 2 · importance
YOUR APP
INSTRUCTION
Does it describe a meaningful life event, goal, relationship, constraint or stable preference?
STATE
observation text, exact_duplicate flag and recent_memories list from the memory policy.
JEV · NOUL
YesNo
Question 3 · novelty
YOUR APP
INSTRUCTION
Does it add a fact absent from recent memories?
STATE
observation text, exact_duplicate flag and recent_memories list from the memory policy.
JEV · NOUL
YesNo
Question 4 · episodic
YOUR APP
INSTRUCTION
Does it describe a particular experience or event?
STATE
observation text, exact_duplicate flag and recent_memories list from the memory policy.
JEV · NOUL
YesNo
Question 5 · semantic
YOUR APP
INSTRUCTION
Does it state a durable fact about an entity or the world?
STATE
observation text, exact_duplicate flag and recent_memories list from the memory policy.
JEV · NOUL
YesNo
App workflow
- Assess new observations
Send the observation, duplicate flag and recent memories through the Jev admission and type question set.
- Build and retrieve memory
Store qualifying observations and use Jev to judge links, retrieval routes and candidate relevance.
- Answer with evidence
Jev checks sufficiency or retrieval stopping; a separate language model synthesizes the final answer from selected memories.
Why it is interesting
It separates frequent bounded memory-management and retrieval judgments from the slower generative answer step, with explicit questions about novelty, relevance, and evidence sufficiency.