North Star — Continuous Working Awareness

An Instar agent should never silently lose track of something that mattered. When an important context surfaces — “we’re testing over Telegram,” “this customer hates jargon,” “the real goal is X, not the bug I’m chasing” — it should be captured automatically, ranked by how much it matters and over what horizon, kept warm while relevant, re-surfaced the moment it matters again, and allowed to fade once it stops. None of it can depend on anyone — user or agent — remembering to do it by hand. Capture is automatic or it doesn’t happen.

That’s the North Star: a continuously-maintained working awareness that keeps the agent grounded no matter when, where, or how it operates.

Note

This page is the digestible version. The full vision doc — with the architecture deep-dive, the honest inventory of what already exists, the live case studies, and the staged evolution path — is the canonical source at docs/NORTH-STAR.md in the repo. It’s a beacon, not a contract; it’s meant to move as the framework learns.

Three facets that are really one thing

“Awareness” has three faces, and they’re the same machine pointed in different directions:

• Awareness of the world — tasks, conversations, goals, constraints. “We’re testing over Telegram right now.”
• Awareness of itself — its own capabilities and features. “I have a Secret Drop feature that’s perfect for this.”
• Awareness of its standards — the principles that guide how it builds. “This feature would violate framework-agnostic.”

An agent that forgets the test surface, an agent that forgets its own feature, and an agent that ships a standards-violating feature are all failing the same way: a relevant context existed and never reached the moment it was needed.

Drift is a context-lifecycle failure

The tempting fix for each kind of drift is a dedicated detector — watch for missing Telegram actions, watch for audience mismatch, watch for goal-drift. That’s whack-a-mole: an ever-growing pile of brittle, single-purpose sentinels, each with its own stale state and its own dismiss-fatigue.

Step back and every one of these is the same failure:

An important context arose, mattered, then aged out of the working set as the task evolved — and nothing pulled it back when it became relevant again.

So the fix isn’t a detector per symptom. It’s one general lifecycle for important contexts: capture → rank → maintain → re-surface → decay. Build that once, and method-drift, audience-drift, and goal-drift all fall out of it for free.

The hierarchy of contexts

Different contexts live on different time horizons, and decay rate is the knob:

• Short-term (this task): “we’re testing over Telegram right now.” High weight, fast decay.
• Medium-term (this project/relationship): “this project auto-publishes on any non-template release note.” Survives sessions, fades over weeks if unused.
• Long-term (identity/values): “Justin wants ELI16, always.” Near-permanent.

Decay is demotion, not deletion — a faded context drops out of the hot set but stays retrievable, and a later reference re-warms it, exactly like human memory.

The architecture

          Live conversation + action stream
                  │ (reads)        │ (reads)
          ┌───────▼──────┐   ┌─────▼──────────┐
          │  LIBRARIAN   │   │  USHER         │
          │ auto-capture │   │ grounding watch│
          └───────┬──────┘   └─────┬──────────┘
                  │ writes          │ queries
          ┌───────▼─────────────────▼──────────┐
          │   WORKING AWARENESS STORE            │
          │   world + capabilities + standards   │
          │   ranked · time-horizoned · decaying │
          └───────┬──────────────────────────────┘
                  │ injection decision (full-context gate)
                  ▼
          Inject into the agent — only when it
          changes what the agent does next.

Two design choices carry the whole thing:

• Signal vs. authority. The Librarian and Usher are cheap and fast, and they only emit signals — “this might matter,” “this might be worth re-surfacing.” A higher-context decision step decides whether to actually inject or stay silent.
• Near-silent by default. A loop that chatters becomes the next thing dismissed 73 times. The Usher injects only when it would change the agent’s next action. Everything else goes to a pull surface, never into the agent’s face.

Not starting from scratch

The North Star isn’t greenfield — it generalizes a loop Instar already prototyped. The in-progress Topic-Intent Layer is almost exactly this capture→rank→decay→inject→gate loop, scoped narrowly to conversational facts: an LLM auto-captures every substantive turn (no manual step), ranks each item with a user-authority clamp, decays them on a grace-period-plus-half-life model, injects a settled-vs-tentative briefing at session start, and gates pre-send drafts.

The genuine gaps are three: generality (capture what I’m doing, not just conversational facts), unification (a dozen single-purpose watchers should feed one ranked working set), and a continuous mid-task injection surface (today injection happens at session-start or pre-send, never in the middle of a task when a faded-but-now-relevant context should return).

Why this is the right north star

The agent’s whole value is coherence — being the same grounded entity whether it’s writing an email, driving a test, or chasing a bug. Drift is the slow leak in that coherence, and Instar has been patching leaks one at a time. The North Star is the decision to build the thing that keeps the agent grounded in general — to make “structure beats willpower” true for attention, not just procedure. Every session the loop runs, the next session starts more grounded than the last.

The Standards Registry is the first tangible artifact of this vision — the normative facet (awareness of standards) made real, ahead of the continuous machinery that will eventually surface it automatically.

---

Disagree with any of this? It’s a beacon, not a contract — it’s supposed to move. The full vision doc has the complete architecture, inventory, and evolution path.