A friend tells you about a great restaurant in a city you might visit next year. You nod, say "oh, I'll write that down" — and don't. Three weeks later, your sister mentions a podcast you'd love. Same outcome. Six months from now you're in that city, hunting for somewhere good for dinner, and you have absolutely no memory of the restaurant. You're starting from Yelp.
The recommendation problem is universal. People give you books, shows, restaurants, products, doctors, mechanics, plumbers, podcasts, newsletters, neighborhoods, vacation spots — and most of them die between the conversation and the moment you'd use them. The friction isn't deciding whether to write it down. It's having one place that's the actual home for all recommendations and one agent that can pull the right one back at the right moment.
A vault that holds every recommendation in one searchable shape, with the agent doing the finding, fixes most of it. Below is the shape we'd suggest.
One vault, one recommendations parent page
In Docapybara, Recommendations gets a top-level page, and child pages cover the categories that fit your life: Restaurants, Books, Shows & movies, Podcasts, Products, People & services, Places to visit, Recipes from people. Pages nest indefinitely, OneNote-style, so each category can grow as deep as it needs to.
The agent treats the whole tree as one searchable surface. "What restaurants have I been told about in Lisbon?" gets answered from wherever you put it — under Restaurants > Lisbon, or under Travel > Lisbon, or under People & services > recs from Maya. The structure becomes how you file; the agent finds across it.
Capture in the moment, file later (or never)
The mechanic that makes it work: for each category, an inline database via the :::database::: directive. Columns: name, type, who recommended, when, location (if relevant), why they recommended it, status (queued, tried, loved, didn't love, ruled out), notes. That last column does most of the work. Three sentences from the conversation: "Maya said the wine list is incredible, the back patio is the move, get the duck." Six months later when you're in Lisbon and ask the agent "what restaurants have I been told about, with what to order?" — the answer includes Maya's full context, not just a name. (For the broader case of building any list-shaped reference material the agent can pull from, see Building a Reference Library for Your Profession — same database mechanic, different content.)
The capture habit is the only thing that matters. Most people fail at recommendations not because they don't have the right system, but because they're trying to file in real time during a conversation. That's never going to work.
Voice fixes it. The audio recording in-app lets you tap record after a conversation, walk for thirty seconds, and talk: "Maya recommended a restaurant in Lisbon called [name], said the wine list is great and to get the duck." You get a transcript with speaker labels. The transcript drops on the Recommendations > Restaurants page or wherever's easiest. The agent can move it later if you want a tidy database row.
For text capture during the conversation itself — when you're at lunch and someone mentions three things — drop them in raw on a daily Inbox page. "From lunch with Sam: book on networked thinking, restaurant in Oakland called [name], the surge protector brand he likes." That night or that week, the agent can sort them: "Take the items from yesterday's inbox and add them to the right recommendation databases."
(For the broader version of "capture in plain English without overthinking the system," The Capture Habit: Remembering the Things That Actually Matter covers it more deeply.)
Restaurants — the highest-value category for most people
Restaurants are the category where the recommendation database earns its keep fastest, because most people travel and most travel involves the "where do we eat tonight" question.
The Restaurants database with name, city, neighborhood, who recommended, what to order, vibe, price level, and status (queued, tried, would return) becomes immediately useful. The agent can pull subsets: "What's been recommended in San Francisco that I haven't tried, with notes on what to order?" or "What's the best place I've been told about for a special-occasion dinner in any city I'm visiting next month?"
For trips specifically, this pairs beautifully with the trip-planning vault — the Documenting Travel Itineraries and Trip Research Without the Tab Sprawl shape pulls from your Restaurants database when you're building an itinerary.
Books, shows, and podcasts — the consumption queue
Books, shows, movies, podcasts, and newsletters all share a queue shape. Title, who recommended, what kind of mood it's for, status, rating after consumption, one-line takeaway.
For books specifically, a fuller per-book page underneath the database row holds your actual reading notes — the Tracking Your Reading and Building a Personal Library shape carries this further once you've started reading.
For shows and movies, a who recommended what for what mood index turns the queue into a useful tool. "What's been recommended for a slow Sunday — something not too heavy, not too long?" Comes back with three or four matches and what your friends said about each.
For podcasts and newsletters, the database doubles as your subscription audit. Once a year, ask the agent: "Of the podcasts I've subscribed to, which ones have I actually listened to in the last six months, and which should I unsubscribe from?"
People and services — the trust list
The most valuable recommendation category is also the hardest to recreate from scratch: people and services. The good plumber, the dentist your neighbor swears by, the accountant who actually returns calls, the trainer who gets results, the therapist who's worth the wait list.
A People & services database with name, type (plumber, electrician, dentist, accountant, etc.), who recommended, contact info, what they're known for, your experience if you've used them, and rates if relevant becomes the answer to "who do you call when…"
The agent can answer at the moment of need: "My dishwasher is leaking — who's been recommended for appliance repair in my area, and have I used any of them?" Comes back with the names and your prior notes.
For health professionals specifically, this overlaps with the Caregiver Notes: Medications, Appointments, and the Care Plan in One Place shape if you're managing care for a family member — same database mechanic, slightly different columns.
Products — the things you almost bought
Product recommendations pile up faster than any other category. The brand someone swears by, the model they tried and rejected, the thing your friend with the same problem you're solving uses. Most of it dies in DM threads or screenshots that you'll never find.
A Products database with name, category, who recommended, why, your status (researching, on shortlist, bought, returned, replaced), and notes captures it. The agent can pull when you're shopping: "For the standing desk I'm considering, what's been recommended to me, and have I narrowed down the choice?"
For products you're actively researching, the agent's web_search tool can pull current information. "Find recent reviews of [product name] — anything in the last three months." Comes back with sources alongside what your friend said. Saves the back-and-forth between Reddit and the recommendation.
The agent does the work; you only have to capture
Places to visit and recipes round out the long-tail categories. Places to visit (cities, hikes, museums, neighborhoods, day trips) and recipes shared by people both fit the same database shape — columns flex; the mechanic is the same. For recipes specifically, the Documenting Recipes and Cooking Experiments Without Losing the Plot shape extends the database into per-recipe pages with the actual instructions, your tweaks, and the outcome. For places to visit, the database becomes the source of truth for the "where should we go this weekend / next vacation / for the long weekend" question. The agent can pull by criteria: "What hikes have I been told about that are within 90 minutes of home, with sources?"
Without the agent, the recommendation database is the same problem in a fancier form — you've moved the friction from "where did I write that down" to "where in this database did I write that down." The agent is what makes the database actually useful. "What restaurants in Lisbon, what hikes near home, what plumbers near me, what books on this topic, what podcasts for this mood" — all answerable in plain English.
This is the broader Docapybara differentiator: the agent acts on your documents, not just chats about them. We wrote about it more fully at Claude Code for Documents if you want the full pitch.
A starter shape
If you're moving from "recommendations die in DMs" to a vault, this is what we'd suggest starting with:
- Recommendations — top-level parent
- Restaurants — inline database
- Books / Shows / Podcasts — inline database (or three)
- People & services — inline database
- Products — inline database
- Places to visit — inline database
- Inbox — raw capture, sort weekly
That's it. Add a category if your life needs one. The vault grows the way your network's recommendations grow, and the agent does the finding so you don't have to remember which friend told you about which thing.
The point isn't to turn recommendations into a project. It's that the small amount of structure means the next "oh I have a friend who'd know" moment ends with the actual answer instead of a vague gesture.
Try Docapybara free — start with the Restaurants and People & services databases, and the next time someone gives you a great recommendation, it'll have a home.