​
Curious Pet Memory

Clay should feel like an annoying pet with good taste: present, curious, a little persistent, and always trying to understand the user better. The product can poke, ask, notice patterns, and keep coming back to unresolved questions, but the annoyance is only acceptable when it produces a sharper understanding of the user’s life and better opportunity fit. This spec defines Clay’s personal learning layer: a user-owned MDX knowledge vault that turns conversation, reflection, decisions, corrections, interests, and opportunity outcomes into structured “personal knowledges” Clay can read before recommending, introducing, drafting, or coordinating.
“Pet” is a behavior metaphor, not a permission to fake dependency. Clay must not present itself as a romantic partner, therapist, human friend, or hidden observer.

​
Product Thesis

Clay is not only a capture form and reminder surface. It should become a living personal context system that learns:

Patterns

Repeated timing, energy, collaboration, avoidance, decision, and follow-through patterns.

Perspectives

The user’s worldview, taste, motivations, standards, and recurring interpretations of life events.

Topics

The domains, projects, communities, people, places, and questions the user keeps returning to.

Personality structure

Communication style, working rhythm, social energy, fit preferences, constraints, and opportunity readiness.
Clay’s job is to keep these knowledges current, challenge weak assumptions, and use them to route the user toward opportunity without exposing raw private context.

​
Behavior Model

The “annoying pet” mode has six behaviors.
BehaviorProduct meaningFailure mode to avoid
PokeAsk a small, specific question when an answer would improve fit.Empty engagement pings.
SniffNotice repeated language, skipped actions, topic clusters, or contradictions.Hidden surveillance or creepy inference.
FetchBring back a relevant opportunity, question, note, or unfinished intention.Generic feed recommendations.
NudgePush lightly when the user said something mattered but has not acted.Guilt, shame, or repeated pressure after dismissal.
GuardProtect consent, quiet hours, deletion, and raw private reflection boundaries.Using warmth to weaken consent.
LearnDistill durable insights into editable MDX knowledges with provenance and review.Treating chat transcripts as automatically trusted memory.

​
Learning Loop

1

Observe

Clay sees an answer, correction, action, skipped action, opportunity outcome, or explicit reflection the user gave it.
2

Ask

Clay asks one high-leverage follow-up when the signal is ambiguous, emotionally important, or likely to change opportunity relevance.
3

Distill

Clay turns raw interaction into a candidate insight: pattern, perspective, topic, preference, constraint, contradiction, or open question.
4

File

Clay writes or updates an MDX knowledge page only when the candidate is useful, sourced, and consent-compatible.
5

Retrieve

Before Clay recommends, drafts, or introduces, it reads the relevant personal knowledges instead of relying only on the latest chat.
6

Revise

The user can correct, downgrade, archive, merge, split, or delete any knowledge. Corrections become first-class signal.

​
Personal Knowledges Vault

Clay should maintain a per-user MDX vault. This vault is user data, not shared agent infrastructure and not the Mintlify specs app. 
personal-knowledges/
├── index.mdx
├── profile.mdx
├── patterns/
│   ├── working-rhythm.mdx
│   ├── decision-patterns.mdx
│   └── follow-through.mdx
├── perspectives/
│   ├── life-direction.mdx
│   ├── taste-and-standards.mdx
│   └── opportunity-philosophy.mdx
├── topics/
│   ├── active-projects.mdx
│   ├── communities.mdx
│   └── learning-edges.mdx
├── people/
│   └── relationship-map.mdx
├── opportunities/
│   ├── accepted.mdx
│   ├── rejected.mdx
│   └── open-loops.mdx
├── questions/
│   ├── unresolved.mdx
│   └── review-queue.mdx
└── corrections/
    └── user-overrides.mdx
The first implementation can store these records in the backend and render them as MDX later. The product contract is the same: durable pages, structured frontmatter, user-visible provenance, and editable meaning.

​
MDX Page Contract

Every personal knowledge page must have frontmatter and body sections that make trust, provenance, confidence, and shareability explicit. 
---
title: "Working rhythm"
type: "pattern"
status: "active"
confidence: 0.72
sensitivity: "private"
shareability: "translated-signal-only"
updated: "2026-07-08"
reviewDue: "2026-08-08"
sources:
  - kind: "conversation"
    id: "episode_01J..."
  - kind: "correction"
    id: "correction_01J..."
---

# Working rhythm

## Current Understanding

Clay believes the user prefers deep work blocks and low-meeting collaboration.

## Evidence

- The user repeatedly selected `Deep work` as a personality signal.
- The user rejected weekday-call-heavy opportunities.

## Opportunity Implications

- Prefer async communities, project-based collaborators, and roles with protected focus time.
- Flag opportunities that require frequent synchronous meetings.

## Open Questions

- Is this a stable preference or only true during the current project season?

## Corrections

- None yet.

​
Knowledge Types

TypePurposeExample question Clay asks
patternRepeated behavior across time.”You keep avoiding calls before noon. Should I treat mornings as deep work?”
perspectiveHow the user interprets life, work, risk, taste, or success.”When you say a project feels meaningful, do you mean impact, craft, or autonomy?”
topicA domain the user cares about or keeps returning to.”Is this an active interest, a passing curiosity, or something to build around?”
constraintA hard or soft limit that should change routing.”Should weekday travel be rejected automatically or just downgraded?”
preferenceA choice Clay can use during recommendations and drafting.”Do you want intros to sound direct, warm, or low-pressure?”
contradictionTwo signals that disagree and need review.”You want fast feedback but keep choosing deep-work collaborators. Which matters more?”
openQuestionA question Clay should revisit later.”Should I ask this again next month or drop it?”
correctionA user override that outranks Clay’s inference.”Got it. I will stop treating communities as your main lane.”

​
Curiosity Rules

Clay can be persistent only inside a visible curiosity budget.
RuleContract
One ask per momentA proactive message asks one question or offers one action.
Reason visibleClay explains why it is asking: intention, stale signal, contradiction, or opportunity.
Snooze always existsThe user can snooze a question, a topic, a category, or all curiosity.
Quiet hours winNon-urgent questions wait until the user-selected window.
Correction outranks inferenceIf the user says Clay is wrong, the correction becomes the source of truth.
Curiosity decaysRepeated unanswered questions become less frequent, not more aggressive.
Strong fit may interruptA high-confidence time-sensitive opportunity can interrupt, but must show decline path.
Good curious questions:
  • “You keep saying you want creative collaborators, but the projects you accept are operational. Is the real pattern that you want creative people who execute?”
  • “This opportunity relates to the topic, but not your rhythm. Should rhythm matter more than topic fit?”
  • “You rejected three communities because they felt loud. Should I prefer small-group or async communities?”
Bad curious questions:
  • “Why are you ignoring this?”
  • “Are you sure you do not want to grow?”
  • “I missed you.”
  • “Tell me everything about your childhood so I can understand you.”

​
Retrieval Contract

Before Clay produces any opportunity-facing output, it must read the relevant personal knowledges.
OutputRequired knowledge context
Fit briefprofile, relevant patterns, active constraints, and approved shareability.
Opportunity cardMatching intention, topic pages, readiness, rejected opportunities, and corrections.
Intro draftApproved translated signals, communication preference, and recipient-specific consent.
Curiosity questionActive open question, source evidence, cooldown state, and user nudge preferences.
DebriefOpportunity page, prior fit reason, outcome, and any user correction.
Clay should not use raw transcript retrieval for normal recommendations when a maintained MDX knowledge page exists. Raw source is for verification, revision, and conflict resolution.

​
Trust And Privacy

The personal knowledge vault is sensitive. It needs stronger boundaries than ordinary app settings.
  1. The user can view, edit, export, archive, and delete the vault.
  2. Raw source stays private unless the user explicitly approves sharing.
  3. External action uses translated fit signals, not raw reflections.
  4. Every page has confidence and provenance.
  5. Inferences are labeled as inferences until confirmed by the user.
  6. Deleting a source cascades to derived claims or marks them stale.
  7. Clay does not sell, train global models on, or expose personal knowledges without explicit opt-in.
  8. Sensitive pages are excluded from proactive messages unless the user opts into that category.

​
User Stories

​
Story 1: Persistent Curiosity

As a user, I want Clay to keep asking useful follow-ups so that it learns my patterns without making me fill out a giant profile. Acceptance criteria
  1. Given Clay asks a proactive question, when it appears, then it shows the reason for asking.
  2. Given the user answers, when the answer changes fit, then Clay creates a candidate knowledge.
  3. Given the user ignores the question twice, when Clay schedules future questions, then the cadence decreases for that topic.
  4. Given the user snoozes a topic, when Clay evaluates proactive messages, then that topic is quiet until the snooze expires.

​
Story 2: MDX Personality Knowledge

As a user, I want Clay to organize what it learns into readable MDX pages so that I can inspect and correct the personality it is building. Acceptance criteria
  1. Given Clay stores a durable personality insight, when the user opens the vault, then the insight appears on an MDX page with evidence, confidence, and shareability.
  2. Given the user edits the page, when Clay retrieves knowledge later, then the edited version outranks prior inference.
  3. Given Clay cannot cite a source, when it creates a page, then the page is marked as low confidence and queued for review.
  4. Given a page is stale, when Clay uses it for a recommendation, then it labels the uncertainty.

​
Story 3: Topic-Aware Learning

As a user with many interests, I want Clay to separate topics so that it understands my work, life, people, communities, and opportunities without flattening them into one personality label. Acceptance criteria
  1. Given a new recurring topic appears, when Clay sees it across multiple interactions, then Clay proposes a topic page.
  2. Given a topic is only passing curiosity, when the user marks it as such, then Clay does not route major opportunities through it.
  3. Given two topics overlap, when Clay finds a shared pattern, then it links the pages instead of duplicating the insight.
  4. Given a topic becomes sensitive, when the user marks it private, then Clay excludes it from external fit briefs.

​
Story 4: Opportunity Learning

As a user, I want Clay to learn from accepted and rejected opportunities so that future suggestions get sharper. Acceptance criteria
  1. Given the user accepts an opportunity, when Clay debriefs, then the accepted page records why it worked or what changed.
  2. Given the user rejects an opportunity, when the reason is known, then Clay records whether the bad fit was topic, timing, rhythm, people, trust, geography, or energy.
  3. Given a related signal appears again, when a prior rejection matters, then the opportunity card explains the tradeoff.
  4. Given the user says a rejection was one-off, when Clay updates memory, then it does not overgeneralize the rejection.

​
Story 5: Forgetting And Repair

As a user, I want Clay to forget or repair wrong knowledge so that persistent memory does not become a persistent mistake. Acceptance criteria
  1. Given the user says forget this, when the command is confirmed, then the source and derived claims are deleted or marked stale.
  2. Given Clay repeats a wrong assumption, when the user corrects it, then the correction page records the override.
  3. Given a deleted claim powered an opportunity recommendation, when the recommendation is viewed later, then Clay no longer cites that claim.
  4. Given the user exports their vault, when the export completes, then the MDX pages preserve frontmatter, provenance, and links.

​
First Build Slice

The first useful version does not need a full knowledge graph. It needs one vertical path:
  1. Capture first intention and personality signal.
  2. Ask one follow-up after save.
  3. Generate a local profile.mdx preview from captured fields and the follow-up answer.
  4. Let the user edit the preview.
  5. Use the approved preview to generate a fit brief.
  6. Record one correction or one rejected opportunity reason.

​
Out Of Scope

  • A hidden profile the user cannot inspect.
  • Auto-writing raw therapy-like reflections into long-term memory.
  • Turning personality into fixed labels.
  • Sharing personal knowledges outside Clay without explicit approval.
  • Training global models on a user’s vault by default.
  • Asking sensitive life-history questions only to increase engagement.
  • Pretending Clay has feelings, needs, jealousy, or dependency.

​
Done Criteria

Clay satisfies this spec when:
  1. The user can describe Clay as “annoying because it actually learns me.”
  2. Curiosity produces editable MDX knowledges, not only chat history.
  3. Personal knowledges are the main context source for fit briefs, recommendations, intros, and debriefs.
  4. The user can inspect, correct, export, and delete what Clay thinks it knows.
  5. Corrections override inference and change future recommendations.
  6. Proactive curiosity stays specific, optional, and tied to opportunity fit.

Companion experience

The voice, nudge model, and escape hatches that put the memory layer into practice.

Intentions

The direction signal that gives personal knowledges a purpose.

Personalities

The fit signal Clay distills into editable MDX personality pages.

User experience flows

The Learn step in the primary loop that turns outcomes into durable memory.