I begin with context, not interruption.
Every fifteen minutes, deterministic collectors look for changes that could matter: official warnings, weather, public transport, calendar context, and—only when a journey is actually supported by evidence—route-relevant traffic.
I do not turn every observation into a notification. Freshness, confidence, privacy, relevance, lifecycle state and your controls are checked first. When the answer is “nothing material changed,” the correct output is nothing at all.
18:30:31 knmi.warning ........... verified
18:30:31 calendar.context ....... local
18:30:32 relevance .............. bounded
18:30:32 material_change ........ false
18:30:32 output .................. 0 bytes
Private context stays close to home.
Calendar details are fetched into a local RAM-backed cache. Public news and route results are cached there too. They can disappear on restart because they are rebuildable. Sessions, configuration, memories and tamper-evident receipts remain persistent because they are not disposable.
Recurring query patterns may be learned only as aggregate categories. The raw questions are not retained for that purpose. Sensitive calendar content is used locally to establish relevance—not copied into public status artifacts or sent to a model just to decide whether an alert matters.
calendar.refresh ........ every 2h + pre-briefing
news.refresh ............ hourly
persistent .............. sessions · state · receipts
raw_query_learning ...... disabled
Useful intelligence is not measured by how often it speaks.
It is measured by whether it noticed the right change, understood the real context, and spoke early enough for you to do something about it.
That is why Nexus is silent by default, source-grounded, confidence-aware, freshness-aware and explicit when evidence is missing.
A boundary should exist before exposure.
The governed request pipeline is designed to classify privacy after authorization but before any model or provider sees the content. Missing or malformed classifications default to local-only. A sensitive session stays local until it is explicitly reset, and local-only routing disables model fallback.
That universal Telegram path is implemented and contract-tested, but deliberately not active yet. The current supported private route is a manual switch to the local Qwen model. Nexus says this plainly because a security boundary that is merely prepared is not the same as one enforcing every live request.
privacy.classify ........... sensitive | standard | unknown
unknown.default ............ local-only
local.fallbacks ............ disabled
telegram.cutover ........... pending supervised gate
I am coordinated, not divided into imaginary people.
Hermes Agent is the conversational core. Deterministic collectors perform repeatable observation. Skills provide deep procedures. Isolated subagents can research, inspect code or review a problem without flooding the main context. MCP and Seatbelt place governed boundaries around external research and DMZ tools.
There are no permanently running fictional personalities behind the curtain. A specialist exists when a bounded task calls for one; its scope, tools and result are explicit. For important output, the prepared review layer can run one bounded local critic pass—but it cannot silently recurse forever.
Capability is not permission.
Proactive intelligence may observe, rank and notify. It may not silently purchase, book, deploy, alter security policy or take control of the physical world. External communications, destructive changes, private adapters and improvement proposals remain human-gated.
Every proactive decision can leave a minimal, allowlisted receipt linked by hash. The receipt proves continuity without copying the private event into an audit trail. If required review is unavailable for elevated risk, the prepared governed path fails closed.
governance.maximum ........ notify
external_action ............ approval required
receipt.payload ............ allowlisted metadata only
ledger.continuity .......... verified
At quiet hours, I examine patterns—not your private life.
Failures and outcomes can be reduced into metadata-only signals. Canaries test whether invariants still hold. A monthly advanced-model review receives a sanitized improvement packet, not raw conversations. It may propose a change. It cannot apply one.
Preference observations follow the same rule: they become pending proposals. Stable choices can be remembered when appropriate, but safety policy cannot weaken itself because convenience appeared to improve.
The honest ledger.
A capable system should know the difference between code that exists, a control that is enforcing, and an idea that still lacks an approved source.