Rowboat – Open‑Source Local‑First AI Coworker with Persistent Memory
Rowboat – Open‑Source Local‑First AI Coworker with Persistent Memory
Rowboat delivers a locally stored, memory‑rich AI coworker
Rowboat is a desktop application that keeps all AI‑generated context on your machine as plain Markdown files, builds a living, backlinked knowledge graph, and provides ready‑made work surfaces (email, notes, browser, code, meeting transcription, and custom apps). This design eliminates the cold‑start problem of most AI tools and gives users full control over their data.
Core architecture: a persistent knowledge graph
Rowboat continuously indexes emails, meetings, Slack, and other assistant conversations into an Obsidian‑style graph where each node is a Markdown file. The graph:
- Accumulates context over time instead of re‑searching on demand.
- Makes relationships explicit; you can inspect backlinks and edit notes directly.
- Remains editable; nothing is hidden inside a proprietary model.
- Lives locally; no cloud lock‑in, easy backup, and full privacy.
"Most AI tools reconstruct context on demand by searching transcripts or documents. Rowboat maintains long‑lived knowledge instead… the result is memory that compounds, rather than retrieval that starts cold every time." – Rowboat README
Built‑in work surfaces act as AI‑augmented tools
Rowboat ships with several surface applications that let the AI act directly on your data:
| Surface | What it does |
|---|---|
| Sorts incoming mail, drafts replies using the full knowledge graph. | |
| Meeting notes | Captures live transcript, summarizes into Markdown, updates the graph. |
| Browser | Isolated web view where the assistant can log in to specific accounts and perform tasks collaboratively. |
| Code mode | Spins up parallel coding agents powered by Claude Code or Codex, with full context from your projects. |
| Apps | Customizable workspaces you can build and share, each with access to all integrations. |
| Background agents | Event‑driven or scheduled scripts that can search the web, call APIs, or write code automatically. |
Extensibility through the Model Context Protocol (MCP)
Rowboat can plug into external services via MCP, a lightweight protocol for passing context between the AI model and tools. Out‑of‑the‑box integrations include:
- Exa for web search
- ElevenLabs for voice output
- Deepgram for voice input
- Composio for dozens of SaaS APIs (Slack, Linear, Jira, GitHub, etc.)
You can also add any custom MCP server, turning Rowboat into a hub for internal tooling.
Bring‑your‑own model, stay model‑agnostic
Rowboat does not lock you into a specific LLM. It works with:
- Local models via Ollama or LM Studio
- Hosted APIs (Claude, OpenAI, Anthropic, etc.)
Switching models is a configuration change; your Markdown vault remains untouched.
Local‑first by design eliminates vendor lock‑in
All data is stored as plain Markdown on your filesystem. This gives you:
- Full inspectability and editability
- Simple backup or deletion at any time
- No proprietary formats or hidden cloud storage
Community feedback highlights strengths and open questions
The Hacker News discussion surfaced several practical viewpoints:
Neozino asked how long development took, indicating curiosity about the effort required for such a system.
TomComb wondered whether the "Agent Apps" become the primary artifact, contrasting Rowboat’s surface‑centric approach with traditional project‑folder plugins.
ActionHank warned about the "asymmetry of effort" where AI tools generate more reading material than they save, a reminder that Rowboat’s memory must be managed wisely.
Danny O'Brien sought collaborative prompting features, asking if multiple users can share a conversation—a potential future extension.
_puk shared a real‑world workflow: pointing Claude at the Rowboat directory to retrieve context, and expressed interest in a plugin‑style architecture for custom formats.
Snootypoot praised the harness as feature‑rich and well‑suited to their needs, confirming the project's practical appeal.
These comments collectively suggest that while Rowboat’s local‑first memory model is compelling, users are looking for:
- Team collaboration (shared sessions, pair‑prompting).
- Memory management tools (pruning, summarization, opinionated retention policies).
- Clear onboarding from existing Claude harnesses or other AI workflows.
Getting started
- Download the latest binary for macOS, Windows, or Linux from the Rowboat website.
- Configure optional services (Google, Deepgram, ElevenLabs, Exa, Composio) by placing API keys in
~/.rowboat/config/*.json. - Launch Rowboat and let it index your existing files; the knowledge graph builds automatically.
- Explore the built‑in surfaces or create custom apps via the MCP integration framework.
Why Rowboat matters
Rowboat demonstrates that an AI coworker can be privacy‑first, extensible, and truly persistent without relying on cloud storage. By treating the knowledge graph as a first‑class citizen, it shifts AI from a stateless query engine to a long‑term partner that remembers, organizes, and acts on your work history. This approach could redefine how developers, knowledge workers, and teams interact with large language models—moving from “search‑then‑answer” to “continuous‑collaboration”.