◉ Continuum

◉ Continuum

Git remembers what you changed.Continuum remembers why.

bunx continuum init && bunx continuum start

Every commit becomes knowledge. Decisions, patterns, hard-won insights — extracted automatically, stored locally, available in every AI tool you use.

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.

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.

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.

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.

Works with everything

MCP is the USB-C of AI tools. One protocol, every tool.

{
  "mcpServers": {
    "continuum": {
      "command": "bunx",
      "args": ["continuum", "--mcp-only"]
    }
  }
}

Claude Code: ~/.claude.json · Cursor: ~/.cursor/mcp.json · Others: check their docs.

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

What you're NOT paying for

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.

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

MCP tools (available in any AI tool)

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.

Security

Configuration

~/.continuum/config.json:

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.

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

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

MIT License · Built by @devjoaocastro

The missing memory layer for software development.