Thinking Space

The problem

Chat with an AI for long enough and the conversation forgets itself. The model has a context window, the user has a life, and the gap between what was said three weeks ago and what the model can see today is the gap between a real thinking partner and a stranger that nods politely. Stuff a transcript into a vector store, fetch the top five hits per turn, and you get an AI that recalls things that sound like the current message but cannot tell which moments the user themselves flagged as important.

The solution

A Next.js App Router app where every chat goes through a single streaming POST. The system prompt is assembled fresh from up to seven sources, memory is retrieved across five purpose-built channels, long histories are hybrid-summarised, the response is streamed, and four to five background extractors fire after the stream closes, themes, memories, people and turn detection all happen without blocking the user's next message.

• Five-channel memory · the system's centre of gravity

  Resonance (what sounds like what they just said), Anchors (what they themselves flagged, turns, engaged reflection cards, completed check-ins), Arc (what else is happening in the story they're in the middle of, weighted by theme momentum), State (what's been on their mind in the last 14 days), Echo (an older thought that rhymes, fired probabilistically 40% of the time so it stays a surprise). One query embedding, reused across channels.

• Theme arc · computed, not predicted

  Every theme carries momentum_score (exponential decay, 14-day half-life), layer_velocity (layers per week over the last 30 days), sentiment_trajectory (flat, rising, falling, oscillating, V-shaped, peak-shaped), co-occurring themes, and arc_computed_at. Recomputed after each new layer. The Arc memory channel reads momentum_score directly to pick which themes are still hot.

• Daily reflection cron with rotation

  Vercel cron runs nightly with a CRON_SECRET bearer token. Per-user interval overrides, 30-day rotation window per theme, 14-day duplicate suppression. Card types: recurring, absent, positive_shift, emotional_pattern, focus_related, the type drives the prompt. Cards are embedded on insert so they're ready for the Anchors channel the moment a user engages.

• Profile that unlocks with use

  20 questions defined as constants. The first two unlock immediately. Each subsequent question unlocks after one qualifying chat (≥4 messages). Most have curated pill options plus free-text. Profile answers are injected into the system prompt only if the user has actually filled any in.

• Voice in and voice out · safely

  whisper-1 transcription with 25 MB max upload, MIME validation with codec params stripped, an explicit allow-set, 10/min rate limit, empty-file guard, and a steering prompt that nudges Whisper away from hallucinating numbers on short clips. tts-1 playback with four voice options (US/UK male/female).

• Stream trailers · UX shortcut, not a second round-trip

  __TS_CHATID__ is sent first so the sidebar shows the new chat instantly, before the model response arrives. __TS_TURN__ markers stream inline so turn pivots render as they happen. __TS_META__ JSON is appended at the end with title, summarisation info and (for opted-in users) the retrieved memory bundle for debug.

• Multi-tenant from day one · not retrofitted

  Every user table has RLS enabled and a "users can read own X" policy. Service-role client only used for genuinely cross-user operations (cron, fire-and-forget extractors). Middleware enforces auth for /chat, /themes and /admin. Admin route additionally checks an ADMIN_EMAILS env allowlist plus a super_admins DB table. createServerSupabaseClient() is the only sanctioned factory.

• lib/constants.ts · 265 lines of explicit thresholds

  Hardcoded model names, magic numbers, .select(*), and "as any" are all banned in code review. Every threshold (memory similarity floor, Echo probability, Echo age cutoff, channel slot counts, momentum half-life) lives in one file. Tuning is a code change in one place, not a hunt.

What it deliberately isn't

A generic chatbot. There's no model picker, no personality selector, no temperature slider. The product is a single AI character with an opinionated memory architecture and an opinionated way of seeing themes. That choice cuts the surface area to what actually matters and lets the engineering go deep on the few things that determine whether the app feels like a real thinking partner.

The outcome

• Memory that has shape. Five channels with explicit jobs replace top-K guesswork. The right kind of memory surfaces for the moment.
• Themes that have arc. Momentum, velocity, sentiment trajectory and co-occurrence, the AI can say "this is the one you've been circling all month" because it can compute it.
• Reflection that rotates. Daily cron with 30-day rotation and 14-day duplicate suppression. Cards arrive when they fit.
• Onboarding that doesn't cliff. Profile unlocks as the user uses the product. Each new question is a small reward, not an entry tax.
• Voice without exposure. Whisper rate-limited, MIME-validated, size-capped. tts-1 keyed to four named voices.
• Multi-tenant ready. RLS on every user table from day one. Going public is a switch, not a rebuild.