How Memory Lane works

An agent that prepares reminiscence sessions, where Qloo decides what culturally fits and code checks every item before it is shown.

From favourites to a session

Before · Favourites
The caregiver
A few public favourites: a singer, a film, a dish, a place. Plus where they grew up (town level) and anything to avoid.
1 · Match them in Qloo
Qloo decides, the caregiver confirms
qloo_describe for singers, films, TV and places; qloo_find_tags for dishes and interests. When there are several matches, the caregiver picks one. Nothing unconfirmed is ever used.
2 · Build the Life Taste Profile
Qloo recommends, code checks
qloo_recommend from the confirmed favourites' Qloo IDs, each domain from its own favourites; places near where they grew up; dishes from Qloo's food tags. When too few films fit their years, it asks Qloo again from the ones that do. Code keeps it to their era, lasting places and real dishes, leaves out the avoid-list and painful subjects, and logs every step, including what it left out.
3 · Plan a session
Code composes from the profile; the AI words the prompts
Six items for about 25 minutes: an opening song, a film or TV memory, food, a place, a choice, a closing song. Gentle prompts that invite, never test. Links to YouTube, Spotify or a map, built in code.
4 · Learn from how it landed
Qloo again, from the reactions
qloo_recommend finds more like what they sang along to, and qloo_rank orders everything still to play against it, so Qloo decides what comes first next time. Anything that upset them never returns and joins the avoid-list. The next plan says what changed, and each item shows the calls behind it.

Every item carries the ID of the Qloo call it came from; the evidence drawer shows that record.
Validators run before anything is shown: Qloo evidence, the avoid-list, wording that talks about groups (never about the person), prompts that aren't memory tests, and only allowed links.
When Qloo can't answer (no key, quota, outage), the app says so. It never fills the gap with made-up items.

The design is “the AI proposes, Qloo decides, code verifies”.

Fills gaps, then Qloo checks. When Qloo leaves a gap (say, no TV from their years), a language model suggests era-typical names. They count only once Qloo has matched and ranked them and their years fit.
Words the prompts, then code checks. It writes warmer prompts for items already in a session; each passes the same checks as the built-in ones, and the avoid-list.
Plays the generic AI. On every demo profile, “Generic AI vs Memory Lane” shows the same model asked without Qloo, and what code counts in each plan.

If the AI is switched off, the app runs on Qloo and code alone, and says so.

Nothing personal reaches Qloo: a guard in front of every call refuses nicknames, contact details, addresses and health terms.
Places stay at town level. Favourites are public cultural names. The nickname stays in Memory Lane.
The AI (OpenAI) gets public facts only, behind the same guard as Qloo: their years, the town, the favourites, the things to avoid (never the reasons) and Qloo's one line about each item. Never the nickname or notes. Its answers are cached without the facts that produced them.
No login: a profile is tied to the device that made it (it is kept on our server). "Delete this profile" removes the profile, its sessions and its Qloo call records; Qloo answers shared by everyone stay cached for up to 8 days. Idle profiles go after 90 days.

Qloo describes what groups of people tend to enjoy, not what one person will like or remember.
Qloo has no year filter. Memory Lane keeps films and TV to the right years in code, and orders singers by the decades Qloo's descriptions mention; it isn't perfect.
Dishes are tags in Qloo, not things with a year or place, so food is lighter than music.
Most places Qloo knows near a town today are new; only lasting landmarks are kept, so some towns get few places.
Reminiscence is an activity, not medical care. Memory Lane makes no medical claims.