Memory Lane

A team of NVIDIA Nemotron models on Nebius Token Factory plans an era-matched week of reminiscence activities for each care-home resident, from just a birth year and one thing they loved.

Comment

Inspiration

About 70% of care-home residents aged 65+ in England live with dementia (Alzheimer's Society). Reminiscence and music modestly improve mood and communication, especially in care homes (Cochrane 2018, 2025), and work best with a person's own years, roughly ages 10 to 30. But activity staff are stretched (96,000 vacancies in England's adult social care, Skills for Care 2025/26), so most residents get the same generic "1950s playlist". We wanted an agent that researches what each person grew up with, makes careful judgement calls, and learns from how each session went.

What it does

Staff type a birth year, a hometown (optional) and one thing the resident loved, e.g. "1946, Liverpool, Cilla Black". In under a minute Memory Lane returns a printable seven-day plan from their own memory years: music, films, telly, a taste of home, armchair travel, stories and radio, and a Sunday sing-along.

  • "Why this will land" for every item, from the Nemotron 3 Ultra planner
  • "Left out for them": picks flagged as likely to upset this person, with reasons
  • Local history hooks: Tavily searches their hometown in their decades, and Nemotron 3 Nano turns results into sourced conversation starters
  • Conversation prompts and follow-ups, matched to dementia stage
  • Word-for-word session scripts from Nemotron 3.5 Lightning in about 2 seconds
  • "How did it go?": staff rate each day, and Nemotron adapts next week with a "What changed" panel
  • Group sessions, model badges showing which Nemotron did each step, and an MCP server

How we built it

Next.js, React and TypeScript on Vercel, with four Nemotron models on Nebius Token Factory (OpenAI-compatible API):

  • Nemotron 3 Ultra, reasoning on: an agentic planner with Qloo tool calls (qloo_search, qloo_more, submit_week)
  • Nemotron 3 Super: writes the week and the group order, and reads feedback (reasoning on)
  • Nemotron 3 Nano: prompts, the notes reader and the Tavily history hooks
  • Nemotron 3.5 Lightning: session scripts

Reasoning is toggled per call with chat_template_kwargs enable_thinking. Each role has a fallback chain and a time budget, with template fallbacks so the app never breaks. Nemotron can only pick Qloo items by alias, so it can't invent titles, and a deterministic safety guard runs on top. Tavily runs in parallel with Qloo, using only town, country and decades. Names never go to Qloo or Tavily.

Measured latency: Ultra planner turn 11.5 s median; Super writing 3.5 s; Nano prompts 3.7 s; Lightning script 2.0 s; a whole week 34 to 44 s.

Challenges

Reasoning tokens consumed max_tokens and left content null. Finding the working reasoning switch (/no_think and reasoning_effort didn't work). Fitting Ultra's 10 to 20 s turns into an interactive app with parallel tools and a live progress card. Keeping an LLM inside the taste graph.

Accomplishments

A real multi-model agent with every model visible in the UI; a closed plan, session, feedback, new week loop; nothing invented and nothing unsafe let through; sourced local history a carer can say out loud.

What we learned

Per-call reasoning control matters more than model choice for latency. Big reasoning models are best kept for the few judgement calls, with small models doing the volume.

What's next

Saved resident profiles, stricter structured outputs, decade-tight Tavily hooks, and pilots with UK care-home activity coordinators.

Demo: https://memory-lane-nemotron-nebius.vercel.app Code: https://github.com/OlsOls4/memory-lane/tree/nebius-nemotron

Built With

Share this project:

Updates

Submission history