91 lines
4.5 KiB
Markdown
91 lines
4.5 KiB
Markdown
# DESIGN_POV.md — strategic notes from the design pass
|
||
|
||
> These are **convictions, not tickets** — observations from living inside Waggle
|
||
> while designing it. They go beyond the brief. One (#1, Memory Trust) is already
|
||
> built into this package as a screen (`screens/memory-trust.html`, doc section
|
||
> "Trust"). The other four are open questions worth a decision before — or shortly
|
||
> after — the refactor ships. Treat this as a starting point for the team, not a
|
||
> spec.
|
||
|
||
---
|
||
|
||
## 1. Memory you can trust · **BUILT — see `memory-trust.html`**
|
||
|
||
**The risk.** A persistent-memory product's #1 churn driver isn't *forgetting* — it's
|
||
**remembering the wrong thing**, acting on something stale, or knowing something the
|
||
user wishes it didn't. The concept (rightly) sells memory as an asset that compounds.
|
||
But the first time Waggle confidently does the wrong thing because of a bad memory,
|
||
trust breaks — and trust is the entire reason someone leaves their working life in the
|
||
hive.
|
||
|
||
**The response (designed).** A *Memory Trust* layer:
|
||
- **Confidence + freshness** on every memory (the agent and the user can both discount
|
||
what's shaky or old instead of acting on it blindly).
|
||
- **Forget** (real removal from recall + from anything said next) and **Correct**
|
||
(inline; dependents re-checked) on any memory.
|
||
- **Stale review** prompts ("this is 6 weeks old — still true?").
|
||
- **"Why did you do that?" trace** — any agent action unfolds into goal → recalled
|
||
memories (with confidence) → checks → action, and the fix (correct/forget the
|
||
offending memory) is one click from the explanation.
|
||
|
||
**Why it's also a moat.** Editable, accountable memory is *harder* to build than
|
||
accumulating memory, and it's exactly what enterprise/KVARK buyers will demand
|
||
(auditability, right-to-correct, EU AI Act alignment). It turns the scariest property
|
||
of the product into its most trustworthy one.
|
||
|
||
---
|
||
|
||
## 2. The cold-start inversion
|
||
|
||
**The problem.** Waggle is **worst on day one** (knows nothing) and best on day 365 —
|
||
the value curve runs backwards exactly when you're trying to convert someone. Import
|
||
helps, but the first week is make-or-break and isn't yet designed as a deliberate arc.
|
||
|
||
**Suggested response.** A "worth it before it's full" thread: honest, inviting empty
|
||
states; and an explicit, **celebrated first aha moment** — the first time Waggle
|
||
surfaces *"I remembered X so you didn't have to,"* made a felt moment rather than a
|
||
silent convenience. Measure time-to-first-recall as a north-star activation metric.
|
||
|
||
---
|
||
|
||
## 3. Voice & lexicon (commit to the metaphor — carefully)
|
||
|
||
The bee system (Waggle / hive / waggle-dance / "while you slept") is charming but
|
||
half-committed; it wobbles between cute and technical. There's a coherent, restrained
|
||
metaphor here (the dance = diffusion is already used well). **Write a short voice &
|
||
lexicon guide:** which bee-words earn their place (hive, waggle-dance), which stay
|
||
plain (memory, agents, connectors), and a hard rule that the enterprise/KVARK surface
|
||
stays sober. Consistency makes it feel intentional instead of themed.
|
||
|
||
---
|
||
|
||
## 4. Who pays for inference?
|
||
|
||
The pricing story ("memory free forever, pay for scale") is clean — but the project's
|
||
**own benchmark admits high token use**, and tokens are the real cost. The unmade
|
||
decision: **BYO-key** (user pays the provider directly — fits local-first, keeps
|
||
margins clean, but adds onboarding friction) **vs. Waggle-metered** (smoother UX, but
|
||
you carry inference cost + need usage caps). This quietly reshapes Billing, Onboarding
|
||
(the model gate), and Usage. Decide it explicitly before the billing flow ships; the
|
||
current design supports either but commits to neither.
|
||
|
||
---
|
||
|
||
## 5. The morning briefing as a *ritual*
|
||
|
||
The habit loop is solid; the most ownable version of the trigger is a **60-second
|
||
briefing you could listen to** — "Waggle reads you your morning." Audio is sticky,
|
||
daily, hands-free, and nobody in the agent category has it. A natural extension of the
|
||
Home cockpit + the "while you slept" summary, and a strong hook for the mobile +
|
||
messaging surfaces on the roadmap.
|
||
|
||
---
|
||
|
||
### How to use this
|
||
- #1 is done — review the screen and wire it to the real memory store
|
||
(confidence/freshness/forget/correct/trace are all backed by data the substrate
|
||
already has or can derive).
|
||
- #2–#5 are decisions for product + design leadership. None blocks the refactor, but
|
||
#4 (inference cost) should be settled before Billing goes live, and #1's trust
|
||
primitives should land early because everything else trades on them.
|