Memento — Google I/O 2026 Hackathon Submission

Memento is a local-first knowledge layer over your email archive, built as a showcase of the Gemini Interactions API. Instead of being yet another email client, it turns years of correspondence into source-attributed, living documents across five dimensions: People, Projects, Newsletters, Concepts, and a unified Dashboard — all powered by streaming, multi-turn Gemini agents.

This submission demonstrates end-to-end use of the Gemini Interactions API: multi-turn agentic loops, server-sent event streaming, tool-call orchestration, and previous_interaction_id-based conversation chaining — all running locally against your own email archive, with zero data sent to a third-party backend.

What it does

• Dashboard — executive brief with an "Ask Memento" chat. Ask about a person, project, or topic and watch the agent call tools, surface entity summaries in a live side panel, and synthesize an answer across your full email history.
• People — relationship wikis for meaningful contacts, resolved across email aliases. An enrich agent writes facets (role, affiliation, expertise) and narrative sections (summary, interaction arc, current status) backed by cited messages.
• Projects — bounded narratives for work/life projects curated via a collector agent chat flow, then compiled section-by-section with citations by the project compile agent.
• Newsletters — auto-detected broadcast sources with Gemini-generated summaries.
• Concepts — opt-in evergreen topics curated and compiled the same way as projects.

Every factual claim traces back to source messages via citations. User edits to generated documents survive re-compiles.

Demo

# 1. Build the Go backend
cd backend && go build -o ../memento ./cmd/memento && cd ..

# 2. One-time setup (runs migrations, prompts for Gemini API key)
./memento init

# 3. Boot the backend
./memento serve --port 8787 &

# 4. Boot the Next.js UI
pnpm install && pnpm run dev

# 5. Open http://localhost:3000

Type a question in the "Ask Memento" box on the dashboard. The Gemini Interactions API streams tool calls and text deltas in real time; the side panel materializes entity cards as the agent surfaces them.

Prerequisites

Runtime:

• Go 1.26+
• Node.js 22+, pnpm
• A Gemini API key — get one free at aistudio.google.com

Email archive (msgvault):

Memento reads from a SQLite archive managed by the msgvault CLI. You need msgvault installed and configured at ~/.msgvault/config.toml pointing at a populated archive for the People, Projects, Newsletters, and Concepts dimensions to populate with real data.

# ~/.msgvault/config.toml
[data]
data_dir = "/absolute/path/to/your/vault"

Useful read-only msgvault commands to verify your setup:

msgvault stats
msgvault list-senders --limit 10 --json
msgvault query "SELECT COUNT(*) FROM v_messages"

Environment

.env at the repo root (created by ./memento init):

GO_BACKEND_URL=http://127.0.0.1:8787
GEMINI_API_KEY=<from Google AI Studio>
INTERNAL_TOKEN=<generated by init>
AGENT_STEP_LIMIT=40
MEMENTO_AGENT_MODEL=gemini-3.5-flash

Tech stack

• Frontend: Next.js 16 (App Router), React 19, TypeScript 5, Tailwind CSS 4, @google/genai 2.6.0 (Gemini Interactions API)
• Backend: Go 1.26, modernc.org/sqlite (pure Go, no CGO), github.com/google/generative-ai-go (newsletter summaries)
• Design system: Archive Atlas — #12362e primary, #faf9f7 background, Source Serif 4 + Inter

Repo layout

src/             Next.js UI + agent runtime
  server/agents/ Gemini Interactions API client, loop, SSE controller
  components/    UI components including agent cards and chat
backend/         Go HTTP server + SQLite migrations
  cmd/memento/   CLI subcommands (init, serve, migrate, refresh, …)
  internal/      msgvault reader, person resolver, dimension stores

Gemini Interactions API

The core of this demo is src/server/agents/loop.ts, which implements the canonical streaming loop:

const stream = await client.interactions.create({
  model: agent.model,
  input: [{ type: "text", text: userMessage }],
  tools: agent.tools,
  system_instruction: agent.system,
  previous_interaction_id: prevId,   // chains turns server-side
  stream: true,
  generation_config: { thinking_level: "high" },
});

for await (const ev of stream) {
  // ev.event_type: "step.start" | "step.delta" | "interaction.completed"
  // Tool calls arrive as steps; results are fed back to continue the loop
}

Each agent (person enrich, project compile, concept compile, Memento router) is a thin wrapper around this loop with its own system prompt and tool set. The Next.js API routes stream SSE events to the browser; the client renders them live.

Architecture Diagram

                ┌────────────────┐
                │ memento-agent  │  ← dashboard hero chat (router)
                └───────┬────────┘
                        │ tool calls
        ┌───────────────┼───────────────────┐
        │               │                   │
┌───────▼──────┐ ┌──────▼───────┐ ┌─────────▼────────┐
│ collector-   │ │ person-agent │ │ project-agent /  │
│ agent        │ │ (facets +    │ │ concept-agent    │
│ (draft flow) │ │  narrative)  │ │ (narratives)     │
└───────┬──────┘ └──────┬───────┘ └─────────┬────────┘
        │               │                   │
        └───────────────┼───────────────────┘
                        ▼
            ┌─────────────────────┐
            │ msgvault (SQLite)   │  ← local, read-only
            │ memento_* tables    │  ← local, Memento-owned
            └─────────────────────┘