The Graph Model Explained
How Atomic tracks every change—and why nothing is ever truly deleted
Understanding the Numbers
When you use Atomic, you'll see output like this:
$ atomic record -m "initial"
[dev 1/XWMUEDIX] initial
1 file changed, +3 vertices, ~0 edges, 21 bytes
$ atomic record -m "edit file"
[dev 2/YRFGB3CG] edit file
1 file changed, +1 vertices, ~1 edges, 30 bytes
What do +3 vertices and ~1 edges mean? Why does creating a file use 3 vertices, but editing it only uses 1?
This guide explains Atomic's graph model in plain terms—and reveals why this design is revolutionary for tracking AI-generated code.
How Atomic Stores Your Files
Unlike Git (which stores snapshots), Atomic stores your files as a directed graph:
- Vertices = chunks of content (pieces of text)
- Edges = connections showing the order of content
Think of it like a chain of puzzle pieces, where each piece is some text and the connections tell you what order to read them in.
FileAdd: Creating a New File (+3 vertices)
When you create a file called hello.txt with content Hello World, Atomic creates 3 vertices:
┌─────────────────────────────────────────────────────────────────────────┐
│ FileAdd: 3 Vertices Created │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ ROOT (virtual) │
│ │ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ NAME VERTEX │ ◄── "hello.txt" (the filename) │
│ │ "hello.txt" │ Lives in the parent directory │
│ └────────┬────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ INODE VERTEX │ ◄── Empty marker (like a file's ID card) │
│ │ (empty) │ Survives renames, identifies the file │
│ └────────┬────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ CONTENT VERTEX │ ◄── "Hello World" (actual file content) │
│ │ "Hello World" │ │
│ └─────────────────┘ │
│ │
│ Stats: +3 vertices, ~0 edges │
│ │
└─────────────────────────────────────────────────────────────────────────┘
Why 3 vertices?
| Vertex | Purpose |
|---|---|
| Name vertex | The filename, stored in the parent directory's graph |
| Inode vertex | An empty marker that identifies this file (survives renames) |
| Content vertex | The actual text content |
Edit/Replacement: Modifying a File (+1 vertex, ~1 edge)
Now let's say you change Hello World to Hello Universe.
Old approach (full replacement): Delete everything, recreate 3 new vertices. Wasteful!
Atomic's approach (Edit/Replacement): Only change what's different:
┌─────────────────────────────────────────────────────────────────────────┐
│ Replacement: 1 Vertex + 1 Edge │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ BEFORE: AFTER: │
│ │
│ ┌─────────────────┐ ┌─────────────────┐ │
│ │ NAME VERTEX │ │ NAME VERTEX │ (unchanged) │
│ │ "hello.txt" │ │ "hello.txt" │ │
│ └────────┬────────┘ └────────┬────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌─────────────────┐ ┌─────────────────┐ │
│ │ INODE VERTEX │ │ INODE VERTEX │ (unchanged) │
│ └────────┬────────┘ └────────┬────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌─────────────────┐ ┌─────────────────┐ │
│ │ "Hello World" │──────────────►│ "Hello World" │ DELETED │
│ │ │ mark as │ ╳╳╳╳╳╳╳╳╳╳╳╳╳ │ (edge mod) │
│ └─────────────────┘ deleted └────────┬────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ "Hello Universe"│ NEW VERTEX │
│ │ │ (+1 vertex) │
│ └─────────────────┘ │
│ │
│ Stats: +1 vertices (new content), ~1 edges (mark old as deleted) │
│ │
└─────────────────────────────────────────────────────────────────────────┘
Why +1 vertex, ~1 edge?
| Operation | What it means |
|---|---|
| +1 vertex | The new content "Hello Universe" is a new chunk |
| ~1 edge | We modified an edge to mark "Hello World" as deleted |
Key Insight
The old content isn't erased—it's marked with a DELETED flag on its edge. The vertex still exists in the graph!
The Diff Algorithm's Role
The diff algorithm (Myers or Patience) compares old vs new content and tells us:
┌─────────────────────────────────────────────────────────────────────────┐
│ Diff Algorithm Output │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ Old Content: New Content: Diff Operations: │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────────────┐ │
│ │ line1 │ │ line1 │ │ KEEP line1 │ │
│ │ line2 │ vs │ MODIFIED │ => │ DELETE line2 │ │
│ │ line3 │ │ line3 │ │ INSERT MODIFIED │ │
│ └──────────────┘ └──────────────┘ │ KEEP line3 │ │
│ └──────────────────────┘ │
│ │
│ DELETE = mark existing vertex's edge as DELETED (~1 edge) │
│ INSERT = create new vertex for new content (+1 vertex) │
│ KEEP = do nothing, content already exists in graph │
│ │
└─────────────────────────────────────────────────────────────────────────┘
Quick Reference: Operation Stats
| Operation | Vertices Added | Edges Modified | Why |
|---|---|---|---|
| FileAdd (new file) | +3 | ~0 | Name + Inode + Content |
| Edit (insert only) | +1 | ~0 | Just new content chunk |
| Edit (delete only) | +0 | ~1 | Mark old content as deleted |
| Replacement (change) | +1 | ~1 | New content + mark old as deleted |
| FileDel (delete file) | +0 | ~1+ | Mark file edges as deleted |
| FileMove (rename) | +1 | ~1 | New name vertex + delete old name edge |
Soft Deletes: The Key to Everything
Here's the crucial insight: content is never truly deleted in Atomic.
When you "delete" something, we just mark an edge with a DELETED flag. The original vertex—and all its metadata—remains in the graph forever.
┌─────────────────────────────────────────────────────────────────────────┐
│ Soft Delete Preserves History │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ Git Model (Hard Delete): Atomic Model (Soft Delete): │
│ │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ "old code" │ │ "old code" │ │
│ │ by: Alice │ ──► GONE! │ by: Alice │ ──► STILL HERE! │
│ │ AI: Claude │ │ AI: Claude │ (just hidden) │
│ └──────────────┘ └──────┬───────┘ │
│ │ DELETED flag │
│ ▼ │
│ ┌──────────────┐ │
│ │ "new code" │ │
│ │ by: Bob │ │
│ │ AI: none │ │
│ └──────────────┘ │
│ │
│ In Git: You lose Alice's In Atomic: Alice's attribution │
│ attribution when Bob rewrites is preserved. We know the NEW │
│ the code. code replaced HER code. │
│ │
└─────────────────────────────────────────────────────────────────────────┘
Why This Matters: AI Attribution
This graph model is revolutionary for tracking AI-generated code.
The Problem with Git
In Git, when someone rewrites AI-generated code, you lose the original attribution:
# Alice uses Claude to generate auth code
git commit -m "Add authentication (AI-generated)"
# Bob rewrites it for security
git commit -m "Fix security issue in auth"
# Now: WHO originally wrote the auth logic?
# Git says: Bob. The AI attribution is buried in history.
The Atomic Solution
Every vertex carries its full provenance—forever:
┌─────────────────────────────────────────────────────────────────────────┐
│ AI Provenance Through Edits │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ Change #1: AI generates function │
│ ┌─────────────────────────────────────────┐ │
│ │ Vertex: "function authenticate() {..." │ │
│ │ Author: Alice │ │
│ │ Provenance: │ │
│ │ - AI: Claude Opus 4 │ │
│ │ - Type: code_generation │ │
│ │ - Tokens: 1,247 │ │
│ │ - Cost: $0.03 │ │
│ └─────────────────────────────────────────┘ │
│ │ │
│ │ Change #2: Human fixes a bug │
│ │ (Replacement: ~1 edge, +1 vertex) │
│ ▼ │
│ ┌─────────────────────────────────────────┐ │
│ │ Vertex: "function authenticate() {..." │ ◄── DELETED but │
│ │ Author: Alice │ PRESERVED! │
│ │ Provenance: Claude Opus 4 │ │
│ └─────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────┐ │
│ │ Vertex: "function authenticate() {..." │ ◄── New vertex │
│ │ Author: Bob │ Human edit │
│ │ Provenance: none (human-written) │ │
│ │ REPLACES: [points to Alice's vertex] │ ◄── Link to original! │
│ └─────────────────────────────────────────┘ │
│ │
│ We can ALWAYS trace back: │
│ "This code was originally AI-generated by Claude for Alice, │
│ then modified by Bob (human) to fix a bug" │
│ │
└─────────────────────────────────────────────────────────────────────────┘
What Metadata is Preserved?
Every vertex in Atomic carries rich metadata:
Vertex {
content: "function validate(input) { ... }",
change_hash: "ABC123...",
author: "Alice <alice@company.com>",
timestamp: "2024-01-25T10:30:00Z",
provenance: Some(Provenance {
vendor: Anthropic,
model: "claude-opus-4",
tool: Cursor,
suggestion_type: CodeGeneration,
tokens: TokenUsage { input: 500, output: 747 },
cost: Cost { amount: 0.03, currency: USD },
session_id: "sess_abc123",
}),
}
This metadata travels with the vertex forever—even when "deleted", the vertex exists and can be queried.
Unrecord: Undo Without Losing History
In Git, undoing a commit either:
git revert— Creates a new commit that undoes changes (loses context)git reset— Erases commits entirely (loses everything)
In Atomic, Unrecord preserves everything:
┌─────────────────────────────────────────────────────────────────────────┐
│ Unrecord: Undo Without Losing History │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ Original State: After Bad Change: After Unrecord: │
│ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ │
│ │ "good" │ │ "good" │ DELETED │ "good" │ ALIVE │
│ │ AI:opus │ │ AI:opus │ │ AI:opus │ │
│ └─────────┘ └────┬────┘ └─────────┘ │
│ │ │
│ ▼ │
│ ┌─────────┐ ┌─────────┐ │
│ │ "bad" │ │ "bad" │ DELETED │
│ │ AI:none │ │ AI:none │ (unrecorded)│
│ └─────────┘ └─────────┘ │
│ │
│ The "good" vertex with AI attribution is RESURRECTED. │
│ The "bad" vertex is now marked as deleted. │
│ ALL METADATA IS PRESERVED - nothing is ever lost! │
│ │
└─────────────────────────────────────────────────────────────────────────┘
Credit Command: Who Wrote What?
The graph model enables powerful attribution queries:
$ atomic credit src/auth.rs
Lines 1-50: Alice (human) - original implementation
Lines 51-80: Claude Opus 4 (AI) - generated for Bob
Lines 81-95: Carol (human) - bug fix
Lines 96-120: Claude Opus 4 (AI) - generated for Bob
└── modified by Dave (human) - security review
AI Summary:
- 45% of current code has AI provenance
- Original AI cost: $0.12 (2,847 tokens)
- AI code modified 2x by humans
This is only possible because we never delete the original vertices.
CRDT Properties: Why Order Doesn't Matter
The graph model gives Atomic CRDT-like merge properties:
- Vertices are immutable — Once created, content never changes (we just add new vertices)
- Deletions are markers — Old content isn't erased, just flagged (can be resurrected!)
- Order is flexible — Edges define order, and independent changes merge automatically
┌─────────────────────────────────────────────────────────────────────────┐
│ Why This Enables Clean Merges │
├─────────────────────────────────────────────────────────────────────────┤
│ │
│ Alice's change: Bob's change: │
│ Edit line 2 Edit line 5 │
│ +1 vertex, ~1 edge +1 vertex, ~1 edge │
│ │ │ │
│ └───────────┬───────────────────┘ │
│ │ │
│ ▼ │
│ Merge Result: │
│ Both changes apply cleanly! │
│ +2 vertices, ~2 edges │
│ │
│ Because the vertices don't overlap, there's no conflict. │
│ Each change adds its own vertex and modifies its own edge. │
│ │
└─────────────────────────────────────────────────────────────────────────┘
Comparison: Git vs Atomic
| Feature | Git | Atomic |
|---|---|---|
| Delete old content | Erased from current tree | Soft-deleted (edge flag) |
| Original metadata | Lost on rewrite | Preserved forever |
| AI attribution | Only in commit message | In every vertex |
| Undo (unrecord) | Reverts lose context | Resurrects original |
| Audit trail | Commit history only | Complete graph traversal |
| Legal compliance | Manual tracking | Built into the model |
Try It: See the Graph in Action
# Create a repository
atomic init my-project && cd my-project
# Create and record a file
echo "Hello World" > hello.txt
atomic add hello.txt
atomic record -m "Add hello"
# Output: +3 vertices, ~0 edges
# Edit the file
echo "Hello Universe" > hello.txt
atomic record -m "Edit hello"
# Output: +1 vertices, ~1 edges
# See what's in the change
atomic change
# Shows: ± hello.txt (+1 vertex, ~1 edge: replace)
Summary
The Graph Model in One Sentence
Atomic stores your code as a graph of immutable vertices connected by edges—"deleting" content just marks edges as deleted, preserving all metadata forever.
This isn't just a technical detail—it's what makes Atomic the system of record for AI-assisted development:
- ✅ Legal compliance: Always prove what was AI-generated
- ✅ Audit trail: Complete history of human vs AI contributions
- ✅ Cost tracking: AI token usage preserved even after rewrites
- ✅ Credit/blame: Know who introduced what, even after refactoring
The Replacement model (+1 vertex, ~1 edge) is fundamental to Atomic's value proposition. By keeping old content instead of erasing it, every piece of code maintains its complete lineage.