Building a Coach That Remembers You
Shipping personalized AI without giving it a script
At Alma, we’re on a mission to build the best AI nutrition coach. But a coach that doesn’t know you is just a chatbot with opinions. Tracking is the foundation, and last year, we hit a wall.
Users loved the natural language experience. But as we scaled, the cracks showed. The same user might say “my usual smoothie” on Monday and “the green one I had yesterday” on Tuesday. They’d track “chicken stir fry” and expect Alma to remember the exact recipe they described three weeks ago.
Our food logging system learned to handle this (we shipped that story here). But knowing what you eat is only half the battle. The other half? A coach who actually remembers your journey.
The Problem with Generic AI Coaching
Before we rebuilt the coaching system, every conversation started from scratch:
Users re-explained their preferences (”I’m vegan, remember?”)
The AI gave generic advice (”aim for 0.8g protein per pound”)
No continuity between conversations
No proactive engagement based on patterns
It felt like meeting a new nutritionist every time you opened the app. Not because the AI was “dumb” - but because we hadn’t given it the right structure to be helpful.
The Core Insight: Map Preferences to Frameworks
Our breakthrough came from nutrition psychology research, not prompt engineering.
We studied evidence-based coaching methodologies:
Precision Nutrition (John Berardi, PhD)
Intuitive Eating (Evelyn Tribole & Elyse Resch)
Motivational Interviewing in Nutrition
Solution-Focused Brief Therapy
Behavioral Science (Katy Milkman)
Then we identified two key axes that determined coaching style:
Nutrition experience: How much do they already know?
Feedback style: How do they want to receive coaching?
This created nine coaching combinations. Each maps to specific research-backed approaches.
Example: Same data, three different voices
User logged 6 out of 7 days this week:
Beginner + Celebratory: “You logged 6 days in a row - that’s building a real habit! What helped you stay consistent?”
Experienced + Tactical: “6/7 days logged. You’re building momentum. One strategy: set a phone reminder for the day you typically miss.”
Knowledgeable + Exploratory: “Your 6/7 logging pattern is solid. Research shows consistency trumps perfection for long-term behavior change. What made the difference between the days you tracked vs the one you didn’t?”
The magic: Users set their preferences once. Every interaction adapts automatically.
Three Coaching Experiences
We didn’t build one monolithic chatbot. We built three specialized modes:
1. Weekly Check-ins (Proactive Coaching)
Every Sunday, the coach reviews your week and sends a personalized check-in. Not generic “great job!” messages - specific observations about YOUR patterns.
It looks at your data, identifies what’s working and what isn’t, sets a measurable focus for the week, and creates an opening message that feels like someone spent time with your data.
User impact: You wake up Sunday to coaching that feels personal, not algorithmic.
2. Nutrition Q&A (On-Demand Intelligence)
When you ask questions, the coach combines:
Your actual meal data
Curated nutrition research
Your stated goals
The result: Evidence-based answers grounded in YOUR context.
Ask “How’s my protein?” and you get your protein trend visualization plus analysis based on your muscle gain goal and eating patterns - not generic advice from a textbook.
3. Natural Conversations (Follow-ups)
The coach remembers the conversation. No re-explaining context.
Sunday: “Your protein was inconsistent this week...” Tuesday: “How’s protein going?” Response: “Since Sunday, you’ve hit your goal 2/2 days. That’s the consistency we’re looking for.”
Memory Without Conversation Logs
Here’s the thing about AI: it doesn’t actually “remember” anything. Context windows don’t span weeks.
Our solution: structured memory.
After every weekly review, the coach saves:
Specific observations (numbers, patterns, dates)
What worked and what didn’t
Recommendations given
Next week, it reads its own notes. Just like a human nutritionist would.
Result: Week-over-week continuity that feels natural.
“Last month I noticed you ate less on weekends. You’ve improved - now only 100 cal below average vs 500 before.”
Show Progress, Not Spinners
We rethought how AI responses should feel.
Traditional approach: User asks → Spinner → Wall of text
Our approach: User asks → “Analyzing your protein intake...” ✓ → Graph appears → Analysis continues
When you ask “How are my macros?”, you see:
Brief context
Graph visualization (appears inline)
Personalized analysis
The system orchestrates UI elements, not just text. When it accesses data, visualizations appear at the right moment in the response.
Users tolerate latency when they see progress. The experience feels interactive, not like waiting for a slow website.
Evidence-Based, Not Generic
Generic AI gives generic nutrition advice. We built something different.
When you ask a nutrition question, the system:
Finds relevant research articles and excerpts
Analyzes your personal data
Synthesizes both into a response
Example: User (vegetarian): “Should I take iron supplements?”
The coach checks your iron intake from food logs, retrieves research on plant-based iron absorption, considers your dietary preferences, and responds with personalized guidance that references both your data and scientific sources.
This isn’t “retrieve and regurgitate” - it’s synthesis.
Shipping It: Test Kitchen to Production
Phase 1: Test Kitchen We dogfooded it ourselves and with early supporters for weeks. Found critical bugs, built confidence in the system, and learned what actually mattered to users.
Phase 2: Fast Feedback Loops Thumbs down feedback went straight to our inbox. Not buried in a dashboard. Bug reports often got fixed same-day.
What we learned: Speed matters more than perfection. Users forgave early issues because we fixed them fast.
What We Learned
1. Personalization ≠ Bigger Models
We spent more time on the coaching methodology framework than model selection. The right structure beats raw intelligence every time.
2. Context Windows ≠ Memory
Real memory is structured, not conversational transcripts. Save observations, not logs.
3. Evidence-Based > Generic
Grounding responses in research frameworks (Precision Nutrition, Intuitive Eating, etc.) made the coaching feel authoritative. Users trust science + data, not “AI opinions.”
4. Show Progress, Not Spinners
Streaming with visible progress increases perceived speed dramatically. Users tolerate latency when they see work happening.
5. Dogfood Relentlessly
Internal use caught most critical issues before they reached users. Build for yourself first.
6. Ship Gradually, Watch Obsessively
Gradual rollout let us catch issues at small scale. Kill switches gave us confidence to move fast.
7. Fail Gracefully
When the system can’t answer perfectly, it acknowledges limitations. Partial answers beat cryptic errors.
What’s Next
We’re live with all premium users. Weekly coaching runs automatically every Sunday. Conversations feel natural and contextual.
Coming improvements:
Proactive pattern detection between check-ins
Multi-modal coaching (image-based feedback)
Deeper goal management capabilities
The vision: A coach that knows you better than you know yourself. Not because it’s “smarter,” but because it pays attention to patterns you miss.
The Takeaway
We didn’t build a chatbot with nutrition opinions. We built a coach that:
Adapts its voice without prompting
Remembers your journey through structured notes
Shows you data inline, not descriptions
Grounds advice in research frameworks
The key insight: Personalization at scale isn’t about prompt engineering or bigger models.
It’s about:
Mapping user preferences to evidence-based frameworks
Building structured memory, not conversation logs
Showing progress and data inline as responses stream
Failing gracefully when uncertain
Users say it feels like having a personal nutritionist. That’s not because we built smarter AI - it’s because we built better architecture around the AI.
The difference between a chatbot and a coach isn’t intelligence. It’s memory, structure, and knowing when to stay silent.
Curious to try this out? Download Alma and try the premium tier.
This piece really made me think about the true depth of context an AI needs to actually feel like a personal coach, just like I wish my smart speaker remembered my favorite Pilates routines without me listing every single sequence.