Ember

A journal that remembers — and notices what you can't.
An AI journaling companion that remembers your inner life and reflects back the patterns you can't see. Built on HydraDB.

▶  Live demo  ·  Quick start  ·  Pitch deck  ·  Design system

---

The one thing to understand

Write a line about your night. Ember writes back — and it remembers: the third Sunday in a row you've dreaded the week, the fact that every time your head felt clearer you'd been running, the friend you keep meaning to call.

Now flip the Memory switch off, and Ember becomes a kind stranger: "Thank you for writing tonight."

The difference between those two replies is the entire product — and it's powered by HydraDB.

Built for the Agents You Love hackathon. The theme is agents people actually want to use, on a memory layer that lets them remember and evolve. Ember is something you'd open every night — and it is useless without its memory. That's exactly the bar.

Why it's different

Most "memory" demos bolt a notes file onto a chatbot; delete the memory and the demo still works. Here, memory is the mechanic:

• 🧠 Memory is the product. Strip it out and Ember is a generic "that sounds hard" bot.
• 👁 It notices what you can't. Recurring themes, the people who matter, open threads — surfaced across weeks.
• ⚡ The payoff is visible in 30 seconds. The Memory ON/OFF toggle shows, side by side, a companion that knows you vs. one meeting you for the first time — while the records stay right there in the panel, untouched.

The 30-second demo

1. The panel — "What Ember remembers" — is already full of a month of entries and patterns, persisted in HydraDB.
2. Click "use tonight's entry" → "Reflect with Ember." It connects tonight to your past: the Sunday dread, the running you've dropped, the call you keep putting off. The "drawing on" chips are the memories HydraDB's semantic search returned.
3. Flip Memory OFF and reflect again on the same entry — a generic, kind-stranger reply, no recall. The records are still in the panel. Ember just stopped reading them.

How it works

 You write  ─▶  Engine (deterministic)  ─▶  HydraDB (semantic memory)
  an entry       • stores the entry           • POST /context/ingest   (persist)
                 • picks what matters           • POST /query           (recall) ◀─┐
                 • builds a safe skeleton                                          │
                        │                                                          │
                        ▼                                                          │
                 Ember's reflection  ◀── "drawing on" these recalled memories ─────┘
                 (an LLM re-voices the skeleton; it never invents — optional)

The rule that makes it bulletproof: the LLM never decides what's remembered. Which memories surface and the reflection's facts are computed deterministically; the model only re-voices an already-correct draft, and falls back to it on any failure. If HydraDB or the LLM is unreachable, Ember still works and the memory beat still lands.

See AGENT.md and specs/ for the architecture and invariants.

🚀 Quick start

npm install
npm run dev          # client on :5173, API on :8787

Open http://localhost:5173. Works with zero keys — local store + deterministic reflection. Add keys in .env (copy .env.example) to light up the full stack:

• HydraDB (memory): HYDRA_API_URL, HYDRA_API_KEY, HYDRA_TENANT. The panel badge flips to HydraDB and entries persist + recall there. Prove it live:

  npm run hydra:check   # creates the tenant, ingests, and queries — prints the proof

• Ember's voice (optional LLM): ANTHROPIC_API_KEY, or Nebius via NEBIUS_API_KEY.

npm run verify        # typecheck → 19 smoke tests → production build

What's actually real

Nothing here is faked or stubbed:

• ✅ typecheck · 19 smoke tests · production build — all green
• ✅ HydraDB is live, not mocked — entries persist via /context/ingest, reflections recall via /query (semantic, ranked by relevance), proven end-to-end with a real key
• ✅ Runs fully offline on a deterministic engine; the LLM and HydraDB only upgrade it

Built with

TypeScript · React · Vite · Node · Express · HydraDB (memory) · Claude / Nebius (optional voice)

Designed as a system (DESIGN.md): editorial-warm — Fraunces for everything you write, Instrument Sans for the interface, one restrained clay accent on warm paper.

How it maps to the prizes

Prize	Why
Best Use of Memory	Memory is the mechanic; HydraDB does persistence and semantic recall.
Best Agent People Love	A warm nightly companion that notices what you can't.
Best Overall	A complete, polished, genuinely useful product about a real human.
Most Creative	Reflective AI with a visible memory ledger and an honest on/off proof.

---

_{MIT licensed · Submission code MEMORY2026 · Built with Claude Code}