Using curie-agent every day

curie-agent is designed to be your daily AI companion — not a tool you launch occasionally. It wakes up before you do, synthesizes what you've been working on, checks your tasks, and sends you a briefing. Throughout the day it's a command away in your terminal, Telegram, or browser. At night it reviews what happened and updates its memory so tomorrow it's a little smarter.

Enable all three cycles with /heartbeat daily 07:15, /heartbeat dreaming 23:01, and /heartbeat enable. That's all it takes to go from occasional tool to daily companion.

When the daily heartbeat fires, curie-agent reads your MEMORY.md, checks your open TODOs, optionally fetches weather, and sends you a concise briefing — on Telegram if configured, or available in the TUI via /heartbeat now.

☀️ Good morning. Wednesday 28 May.

📋 Active tasks (3):
  • Ship curie-agent wiki engine [in_progress]
  • Review PR #47 — daemon auth fix [todo]
  • Weekly team sync at 14:00 [pending]

🧠 Memory: Last session — fixed path guard on Windows.
   Context fill peaked at 68% before compact.

🌤 18°C, partly cloudy in Warsaw.

Type /todo list for the full task board.

Setup

Choose the interface that fits your current context. All three connect to the same underlying session and memory.

Subagents are in-process TurnLoop instances that run concurrently alongside your main session. Each subagent has its own conversation, tool access, and approval mode — you can keep working while they execute tasks in the background.

Spawn a subagent

Monitor progress

Typical use: run /agent --mode auto "write unit tests for utils.ts" then switch back to your main session for other work. The agent tab lights up when complete.

The deep-research skill activates automatically when you ask for structured research, or when your prompt contains terms like "deep research", "research report", or "compare X and Y in depth". It uses a four-phase methodology that goes far beyond a surface-level web search.

Do a deep research on SQLite-vec vs pgvector for a
local-first vector search use case. I need a comparison
table and a recommendation for my Node.js project.

The planning skill activates when you ask for a plan, roadmap, schedule, or timeline. Before generating anything, it gathers your goals, deadline, resources, and constraints — then selects the right planning mode from seven options.

Each mode comes with a different output format: roadmaps produce phased milestones, timelines work backwards from deadlines, habit plans build sustainable systems. The skill uses prioritization frameworks (Impact × Urgency, RICE) and adds 30–50% time buffer by default.

Plan the launch of my open-source CLI tool. I have 3 months,
working part-time (~15h/week). Goal: 100 GitHub stars and
first paying user in month 3. I have a working MVP but no
docs, website, or marketing presence yet.

curie-agent maintains a three-layer memory system. Each layer is a Markdown file you can read and edit directly. The heartbeat cycles keep them up to date automatically.

When a topic in MEMORY.md grows beyond 10–15 bullets, the agent extracts it into a dedicated file (e.g. memory/projects.md) to keep the main file scannable. Run /memory status to see current file sizes.

Once configured, your agent is fully reachable via Telegram from any device. No port forwarding, no SSH — just your existing Telegram client.

Quick setup (5 steps)

1. Create a bot via @BotFather on Telegram → copy the token
2. Run /channels set-bot-token <token> in curie-agent
3. Message your new bot on Telegram — it will show your user ID
4. Run /channels set-user-id <your-id> and /channels set-chat-id <chat-id>
5. Run /heartbeat enable — you're connected

The wiki is a persistent, compounding knowledge base distinct from MEMORY.md which is agent-authored — facts and heuristics the agent notes about your work. The wiki is source-driven: you feed it articles, PDFs, and URLs, and the agent writes structured pages you can query later. Both live in ~/.curie-agent/ and grow over time.

Start the wiki

This creates ~/.curie-agent/wiki/ with the directory structure, WIKI.md schema, an empty index, and a blank log.

Three workflows

Wiki directory structure

~/.curie-agent/wiki/
├── WIKI.md          # schema + Ingest/Query/Lint procedures
├── index.md         # category-organized content catalog
├── log.md           # append-only chronological activity log
├── raw/             # immutable source files (agent reads, never edits)
│   └── assets/      #   downloaded images and attachments
└── pages/
    ├── entities/    #   one page per person, org, or product
    ├── concepts/    #   one page per topic or idea
    └── summaries/   #   one summary page per ingested source

Power tip: chain with deep-research

Run /deep-research on a topic first to find the best sources. Then ingest them with curie-agent wiki ingest. Your wiki compounds every research session — next time you ask the same question, the answer is already compiled.