Continuum
Automatically extracts architectural decisions, patterns, and insights from Git commits to build a local, structured project memory. It exposes this living context to AI tools via MCP, allowing them to understand the historical reasoning and evolution behind your codebase.
README
<h1 align="center">◉ Continuum</h1>
<p align="center"> <strong>Git remembers what you changed.<br/>Continuum remembers why.</strong> </p>
<p align="center"> <a href="https://github.com/devjoaocastro/continuum/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-MIT-00ffaa?style=flat-square&labelColor=222" alt="License"/></a> <img src="https://img.shields.io/badge/runtime-Bun-00ffaa?style=flat-square&labelColor=222" alt="Bun"/> <img src="https://img.shields.io/badge/protocol-MCP-00ffaa?style=flat-square&labelColor=222" alt="MCP"/> <img src="https://img.shields.io/badge/storage-SQLite-00ffaa?style=flat-square&labelColor=222" alt="SQLite"/> <img src="https://img.shields.io/badge/cloud-none-00ffaa?style=flat-square&labelColor=222" alt="No cloud"/> </p>
<br/>
<p align="center"> <code>bunx continuum init && bunx continuum start</code> </p>
<p align="center"> Every commit becomes knowledge. Decisions, patterns, hard-won insights —<br/> extracted automatically, stored locally, available in every AI tool you use. </p>
<br/>
<br/>
The problem
Your commit says fix: queue deadlock. But the real context was:
Spent 3 hours debugging a Bun ReadableStream deadlock. Async spawn causes freeze. Must spawn synchronously in start(). The fix is non-obvious.
That knowledge is gone. Lost in a diff nobody will read again.
Open a project in 6 months. Your AI asks "why is there a serial queue here?" — and neither of you remembers.
Continuum fixes this.
<br/>
What happens when you commit
$ git commit -m "fix: replace Docker with E2B microVMs"
◉ myproject — 1 new commit(s)
09:41:22 a1b2c3d fix: replace Docker with E2B microVMs ...✓ 3 memories
[decision] Replaced Docker containers with E2B Firecracker microVMs —
Docker cold start was 4-6s blocking UX, E2B boots in 400ms.
10x improvement. Trade-off: E2B is a paid service.
[pattern] Sandbox execution: create → execute → read → destroy.
Always set timeout (30s). Never reuse across requests.
tags: docker, e2b, sandbox, performance
sentiment: positive (improvement)
Every commit becomes structured knowledge. Why, not just what.
<br/>
Your project's brain
After a few weeks, run continuum snapshot:
# myproject — Living Context
## Architecture
- Hono on Cloudflare Workers (not Express — needs edge-compatible runtime)
- D1 SQLite (no RETURNING clause — pattern: INSERT then SELECT)
- E2B Firecracker microVMs (replaced Docker, 10x faster boot)
## Hard-won knowledge
- Bun ReadableStream requires sync spawn in start() — async causes deadlock
(spent 3h debugging, the fix is non-obvious)
- D1 writes serialize through primary region — always batch
- Worker bundle limit 10MB — audit deps before adding anything
## Patterns that work here
- Serial queue for Claude CLI spawns (concurrent = race conditions)
- AES-256-GCM for OAuth tokens (PBKDF2 100K iterations)
- Zod validation at every API boundary, no exceptions
Open this project in 2 years. Your AI already knows everything.
<br/>
<br/>
Intelligence — not just storage
Continuum doesn't just record. It learns.
Temporal awareness
Memories decay over time. A decision from yesterday matters more than one from 6 months ago. But if the same pattern appears in 5 different commits — it gets reinforced. Proven knowledge rises. Noise fades.
Evolution tracking
You switched from Docker to E2B? Continuum doesn't keep two separate facts. It knows Docker was superseded by E2B. Your context stays clean, not cluttered with outdated decisions.
Cross-project knowledge
Working on auth in project B? Continuum knows you solved a Safari ITP cookie issue in project A. It surfaces that knowledge automatically — even though you forgot about it.
Developer DNA
Your aggregate profile across all projects. Technologies you've mastered, patterns you always follow, your decision-making style. A living portrait of how you build software.
$ continuum search "rate limiting"
Found in 3 projects:
◉ api-gateway [pattern] Token bucket with Redis Lua scripts — atomic operations
◉ chat-service [decision] Rate limit at edge, not app layer — 10x fewer requests hit origin
◉ webhook-proxy [gotcha] Stripe webhook retries bypass rate limits — whitelist by signature
Your entire development career, searchable.
<br/>
<br/>
Works with everything
MCP is the USB-C of AI tools. One protocol, every tool.
| Tool | Setup |
|---|---|
| Claude Code | Automatic on init |
| Cursor | Automatic on init |
| Cline / RooCline | 2 lines of JSON |
| Continue.dev | 2 lines of JSON |
| Windsurf | 2 lines of JSON |
| Claude Desktop | 2 lines of JSON |
| Zed | 2 lines of JSON |
<details> <summary><strong>Manual MCP setup</strong></summary>
{
"mcpServers": {
"continuum": {
"command": "bunx",
"args": ["continuum", "--mcp-only"]
}
}
}
Claude Code: ~/.claude.json · Cursor: ~/.cursor/mcp.json · Others: check their docs.
</details>
<br/>
<br/>
How it works
git commit
│
▼
┌──────────────────────────────────────────────────┐
│ Continuum daemon (FSEvents, real-time) │
├──────────────────────────────────────────────────┤
│ │
│ 1. Read diff (secrets auto-redacted) │
│ 2. Claude CLI extracts decisions + patterns │
│ 3. Tag with technologies and sentiment │
│ 4. Detect evolution (reinforced / superseded) │
│ 5. Store in local SQLite with importance score │
│ 6. Expose via MCP to all AI tools │
│ │
└──────────────────────────────────────────────────┘
│
▼
Every AI tool you use knows your project history
<br/>
What you're NOT paying for
| Cloud | None. ~/.continuum/ on your machine. |
| API keys | None. Uses claude -p from your existing CLI. |
| Subscriptions | None. Zero cost beyond what you already have. |
| Data leaving your machine | Never. Unless you opt into git sync. |
<br/>
<br/>
Quick start
Requires: Bun + Claude Code CLI
bunx continuum init # Detect projects, configure AI tools
bunx continuum start # Start daemon — watches commits in real-time
That's it. Make commits normally. Continuum does the rest.
<br/>
Commands
continuum init # Setup — detect projects, configure AI tools
continuum start # Start daemon + MCP server (port 3100)
continuum status # Show projects and memory count
continuum snapshot [project] # Generate CONTINUUM.md — your project's brain
continuum add <project> <text> # Manually save a decision or insight
continuum sync init # Setup cross-device sync (private GitHub repo)
continuum sync push # Push memories to GitHub
continuum sync pull # Pull memories from another machine
<br/>
MCP tools (available in any AI tool)
| Tool | Description |
|---|---|
get_context |
Load project memories — call at session start |
search_context |
Search through all memories with TF-IDF ranking |
add_memory |
Save a decision or insight for future sessions |
list_projects |
List tracked projects with memory counts |
cross_project_insights |
Get relevant knowledge from other projects |
developer_dna |
Your developer profile — tech stack, patterns, style |
memory_timeline |
Chronological knowledge evolution |
<br/>
<br/>
Cross-device sync
Memories travel between machines via a private GitHub repo. No cloud.
continuum sync init # Creates private repo: <you>/continuum-memories
continuum sync push # Export & push
continuum sync pull # Pull on another machine
Auto-sync: set "autoSync": true in ~/.continuum/config.json.
<br/>
Security
| Threat | Protection |
|---|---|
| Secrets in diffs | 20+ patterns auto-redact API keys, tokens, passwords |
| Sensitive files | .env, .pem, .key, *secret* — never read |
| Data exfiltration | Everything stays in ~/.continuum/. No network calls. |
| Supply chain | Zero runtime dependencies. Pure Bun + SQLite. |
<br/>
Configuration
~/.continuum/config.json:
| Setting | Default | Description |
|---|---|---|
projects |
[] |
Git repos to watch |
port |
3100 |
MCP server port |
model |
claude-haiku-4-5-20251001 |
Model for extraction |
ignore |
.env, *.pem... |
Patterns to skip |
sync.autoSync |
false |
Auto-push after extraction |
<br/>
<br/>
The bigger picture
This isn't just a context tool.
It's a permanent, searchable record of your entire development career.
Every project. Every hard decision. Every pattern that worked. Every bug that took 3 days to find. Knowledge that evolves, gets reinforced, supersedes itself — just like your thinking does.
In 5 years, you'll have a memory of everything you built. Not a portfolio. A memory.
<br/>
Roadmap
- [ ] sqlite-vec — semantic vector search
- [ ] Knowledge graph — entity relationships across projects
- [ ] Career stats — "you solved 73 race conditions, 41 auth systems"
- [ ] More backends — Gemini CLI, Ollama, local models
- [ ] Team sync — shared knowledge across team members
- [ ] Menu bar app — native macOS/Windows UI
- [ ] VS Code extension — inline memory annotations
<br/>
Contributing
The core is ~1,800 lines of TypeScript. Read it in an afternoon.
git clone https://github.com/devjoaocastro/continuum
cd continuum && bun install && bun dev
<br/>
<p align="center"> <strong>MIT License</strong> · Built by <a href="https://github.com/devjoaocastro">@devjoaocastro</a> </p> <p align="center"> <em>The missing memory layer for software development.</em> </p>
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。