How OiMy thinks,
remembers, and responds

What it carries: your message text, your user ID (to load your family profile), the channel source (Telegram/WhatsApp/iOS), and a timestamp. Nothing else — no interpretation yet.

How it works (two passes):

Pass 1 — Rule engine (instant): Keyword patterns decide obvious cases. "remind me to" → oimy-reminders. "add eggs to" → shopping-list. If an active conversation thread exists (you were just talking about school), it also considers continuing that thread.

Pass 2 — LLM classifier (if needed): Ambiguous messages go to a small, fast model. It reads a 296-line disambiguation guide plus a directory of all 89 skills and their trigger descriptions. Picks one skill ID. Returns it as a single token in ~80ms.

Disambiguation rules example: "set up a daily check-in for my dad" → the rules say if it mentions scheduling + elder, pick elder-checkin, not elder-care. These rules were hand-tuned for the exact failure cases found in testing.

Three sources of context (injected every turn):

① Entity Facts (Tier 1 — always injected): Names, ages, relationships, medical conditions, allergies, school schedules. These are permanent facts extracted from past conversations. Example: "Mia (daughter, 2y), Pablo (son, 5y), husband travels for work."

② Psych Profile: Communication style (direct vs. venting-first), engagement level, parenting philosophy, emotional state patterns. Determined from conversation history using a personality inference model.

③ Active Loops (up to 2 injected): OiMy tracks 84 ongoing "life situations" — things like sleep regression, school transition, aging parent, fitness goal, or relationship stress. If a loop is active, a one-line context note is added ("Mia is going through sleep regression — do not offer new sleep strategies if already discussed").

④ Honcho Memory: Relevant past messages retrieved from a cloud conversation store using semantic similarity — so OiMy remembers what you said 3 weeks ago if it's relevant now.

Four workflow types:

🗣 Coach — Conversation-first. Asks clarifying questions, validates feelings, builds toward a plan. Used for parenting advice, life coaching, emotional support, relationship guidance. (~40 skills)

⚡ Action — Execute immediately. Parses structured data (times, items, names), calls APIs, confirms result. Used for reminders, grocery lists, appointments, medication tracking. (~30 skills)

🏥 Expert — Evidence-based, careful, non-prescriptive. Provides information with clear boundaries ("see a doctor for diagnosis"). Used for medical symptoms, legal questions, financial decisions. (~10 skills)

💬 Chat — General conversation fallback. Warm, human, no specific structure. Used when nothing else matches clearly.

What the model receives (assembled prompt):
System role + skill instructions → family context block (Tier 1 facts) → psych profile → active loop notes → your message history (last few turns) → your current message.

What it must do: Follow the skill's workflow type, honor communication style preferences (no unsolicited advice if the user is venting), stay within the skill's topic boundaries, and sound like a wise friend — not a chatbot.

Fallback chain: If Grok 4.3 is unavailable or times out (8s budget), Claude Haiku 4.5 takes over automatically.

What gets checked:

• Unsolicited advice: Did the model give advice when the user was just venting? Flag and soften.
• Overpromising: Did the model claim to do something it can't ("I'll book that now")? Remove capability inflation.
• Safety signals: Any indicators of self-harm, domestic violence, or child crisis? Escalate to emergency protocols.
• Null action failures: Did a reminder get saved without a time? Ask the user instead of saving "None".
• Tone mismatch: Did the response feel robotic or list-heavy when the user needed warmth?

Three memory stores updated:

① Entity Facts (SQLite): New names, ages, health conditions, or preferences mentioned this turn get extracted and stored. Confidence scores track reliability. Old facts get a time-based decay (more recent = higher confidence).

② Loop State (SQLite): If a life-situation loop was active, the system records what was addressed. Loops can go dormant (after their natural phase ends) and reactivate if the situation returns.

③ Honcho Conversation Store: The turn is saved to Honcho's cloud memory, making it retrievable in future sessions via semantic search. OiMy never forgets — it just prioritizes what's most relevant.

What makes it feel different from a chatbot: OiMy called you by name, referenced your daughter Mia specifically, didn't give you a 5-point listicle when you needed empathy, and remembered that you've been stressed about sleep regression for 3 weeks. None of that required you to repeat yourself.