Memory OS

A personal life-logging PWA. You brain-dump your day to ChatGPT, paste the result into the app, and Gemini normalizes it into 20 structured dimensions — then Memory OS turns that into dashboards, training history, per-exercise progress, and AI insights.

The core idea: talk once → paste once → get a queryable, visual record of your life.

You ──"log: ..."──▶ ChatGPT ──RAW/NARRATIVE/EXTRACTED──▶ Memory OS /add
                                                            │
                                          Gemini normalizes the EXTRACTED block
                                                            │
                              entries + daily_aggregates (Supabase + pgvector)
                                                            │
                     Dashboard · Training · Progress · Insights · Ask

---

1. Set up your extraction assistant (do this first)

The app expects pasted text in a specific 3-section format. A custom instruction in an LLM produces it. Pick whichever you prefer — both work the same way:

	ChatGPT	Google AI Studio
Best for	Familiar UI, great voice input on mobile	Free premium model (Gemini 2.5 Pro), huge instruction limit
Instruction size	1,500 chars (Custom Instructions) or ~8k (a GPT/Project)	Very large — the full Rich instruction fits easily
Use the	Compact version (or Rich in a Project/GPT)	Rich version

➡️ Full copy-paste prompts + step-by-step setup: chatgpt-custom-instruction.md. In short:

Option A — ChatGPT: Customize ChatGPT (or create a GPT/Project) → paste the instruction → save.

Option B — Google AI Studio (free, more powerful):

1. Open aistudio.google.com → sign in with Google → Create Prompt (Chat).
2. Pick model Gemini 2.5 Pro; open System instructions and paste the Rich instruction.
3. (Optional) set Temperature ~0.3 for consistency, then Save the prompt to reuse it.

Either way: replace 83 in the workout formula with your bodyweight (kg). Then any message starting with log: returns a RAW / NARRATIVE / EXTRACTED block.

Both are just stage 1 (extraction + analysis). The app still uses its own Gemini step to normalize that into the schema before saving — so your choice here only affects extraction quality.

2. Daily flow

1. Talk: log: woke 7:30, gym chest+back, incline press 12.5x15 / 17.5x9 / 20x5, 300g chicken + curd, stress 6, slept 6h
2. Copy the assistant's full reply.
3. Add Entry in the app → pick the day it's for (Today / Yesterday / any date) → paste → Save.
4. Review it in Dashboard, Training, Progress, and Insights.

Missed half a day? Later, pick the same date and paste the rest — entries stack and daily totals add up automatically.

---

Features

Page	What it shows
Dashboard	Today snapshot (calories vs TDEE, protein, sleep, mood, training, energy balance) + Week/Month/Quarter trend charts.
Training	Every day's full detail — each exercise's sets (weight × reps, RPE, assist) and the day's diet + totals.
Progress	Pick an exercise → date-wise weight history, best lift, estimated 1RM, and a progression chart.
Insights	Habits & streaks (consistency heatmap), Patterns (auto-correlations across sleep/mood/calories/etc.), Mind (stress/anxiety/focus, emotions, self-talk), and an AI weekly/monthly review.
Ask	Ask questions about your data (RAG over your entries).
Audit	Review items the normalizer flagged as guessed / ambiguous.
Labs	Upload lab reports; Gemini extracts and analyzes markers.
Profile	Onboarding wizard + ongoing progressive profile prompts.

---

Local development

Package manager: pnpm.

pnpm install
pnpm dev      # http://localhost:3000
pnpm build    # production build
pnpm start    # run the production build
pnpm lint     # ESLint

Type errors don't fail the build (next.config.mjs sets ignoreBuildErrors). Run pnpm exec tsc --noEmit to type-check.

Environment (.env.local)

NEXT_PUBLIC_GEMINI_API_KEY=...      # Google Gemini key (used server-side too)
NEXT_PUBLIC_SUPABASE_URL=...
NEXT_PUBLIC_SUPABASE_ANON_KEY=...
SUPABASE_SERVICE_ROLE_KEY=...       # server-only

Database (Supabase)

Enable the vector and uuid-ossp extensions, then paste the SQL files from supabase/sql/ into the Supabase SQL editor in this order (all idempotent):

1. base-schema.sql — the 10 core tables.
2. all-migrations.sql — extra columns, the entries.embedding index, and the match_entries RPC.
3. auth-rls-migration.sql — auth linkage + strict row-level security. Run after deploying the app.
4. data-backfill.sql — optional one-time link of a pre-auth profile (edit legacy_email).
5. cron.sql — optional pg_cron schedules for the Edge Functions.

Test the AI pipeline

export $(grep NEXT_PUBLIC_GEMINI_API_KEY .env.local | xargs)
npx tsx scripts/test-flow.ts example-text/22may.txt

Runs the real normalizer + sanitizer against a sample paste and prints a gap analysis.

---

Tech

Next.js 15 (App Router) · React · Tailwind v4 · shadcn/ui · Supabase (Postgres + pgvector + Auth/RLS) · Google Gemini (gemini-2.5-flash-lite for normalization, text-embedding-004 for embeddings) · recharts.

For architecture and conventions, see CLAUDE.md.