Owner direction, Pavel Hegler, 2026-09-25 (verbatim in agent-office docs/company/owner-direction/2026-09-25-agency-vision-and-the-knowledge-base.md): we want all the hooks where Jev is deciding whether we write in backant memory or not; eventually we will do this as part of backant memory so that we have jevels in backant memory that enable Jev to make smarter decisions about what goes into memory based on our memory infrastructure. Keep it lean, really structured, the knowledge base but also the memories.
What this asks for, as a design to propose before building:
- A jevel shipped with backant-memory (like agent-office's jevels/, protocol 2 through the jevelry CLI) that answers, for a candidate write: is this durable knowledge or noise, is it already held (near-duplicate against recall), does it belong in STM or LTM, and which existing entry it supersedes or refines.
- memory_write, memory_write_episode and the hook writers (session summary, structural adoption hooks) consult it when a TypeSafe key is present, in the same off/shadow/live modes as agent-office's reflex layer, with today's rule as the fall-back; in live a no drops or demotes the write and the decision is recorded.
- Recall traces already exist; the decisions table for these gates should sit beside them so the tuning rule can judge Jev against the human's later reinforce/demote.
- Lean by design: the jevel's own state stays small (the candidate, the top recall hits, the namespace's counts), and the decisions themselves are pruned with the entries they concern.
Filed as an enhancement for the roadmap; not for immediate implementation. Related: agent-office reflex layer spec 2026-09-23, office_idle decision point task 2026-09-25.
Owner direction, Pavel Hegler, 2026-09-25 (verbatim in agent-office docs/company/owner-direction/2026-09-25-agency-vision-and-the-knowledge-base.md): we want all the hooks where Jev is deciding whether we write in backant memory or not; eventually we will do this as part of backant memory so that we have jevels in backant memory that enable Jev to make smarter decisions about what goes into memory based on our memory infrastructure. Keep it lean, really structured, the knowledge base but also the memories.
What this asks for, as a design to propose before building:
Filed as an enhancement for the roadmap; not for immediate implementation. Related: agent-office reflex layer spec 2026-09-23, office_idle decision point task 2026-09-25.