Skip to content

@akshay_pachaar writes that agent memory pipelines store too much, and that a decision model like Jev is useful when the app has its own rules for what may enter long-term memory.

Guides

Akshay 🚀 @akshay_pachaar

EN

  • 6.7k views
  • 59 likes
  • 103 bookmarks
  • 6 reposts
Poster from Akshay 🚀 (@akshay_pachaar)

About this case

Guide. Akshay (@akshay_pachaar) says an extractor proposes preferences, facts and procedures from conversations, tool results and task outcomes. A travel assistant should not store "I usually prefer an aisle seat," "for this flight, book a window seat," and a 25-minute delay the same way: the aisle preference can help later bookings, the window request belongs to one flight, and the delay is useless once the flight is over. He says not to add a decision model only to repeat what a strong memory system already does. It becomes useful for app rules: only a preference the user stated or confirmed becomes an account-level memory, a preference inferred from one booking stays on that trip, passport numbers and payment details never enter general memory, and ambiguous candidates stay in the session until the user confirms them. Typed questions cover whether the preference was explicit, whether it applies past this trip, what retention scope it should have, and whether reuse on a later booking would create a real risk. Code then applies thresholds. Approved candidates can go to Graphiti as episodes: https://github.com/getzep/graphiti. If Graphiti's current ingestion already meets the app, he says a decision model is unnecessary. He points to the quoted article, titled "Jev Clearly Explained." Video attached.

Original posthttps://x.com/akshay_pachaar/status/2120205022046965836

Related cases