What is CVC? The Time Machine Your AI Agent Has Been Missing

The Problem: AI Agents Have Amnesia

If you've used any AI coding assistant — Claude Code, Cursor, Copilot, Aider — you've experienced this: the AI is brilliant for about 20 minutes, then it starts forgetting what it already fixed, contradicts its own plan, and loops on the same error forever.

This isn't a bug. It's context rot — a fundamental limitation of how Large Language Models work. Research shows that after ~60% context window utilization, LLM reasoning quality drops off a cliff. Bigger context windows don't fix this. They just give the problem more room to spread.

Every AI coding tool today treats conversations as disposable: when you restart, the AI forgets everything. When it gets stuck in a loop, your only option is to start over. When it crashes after 2 hours of work? Gone.

The Solution: CVC — Cognitive Version Control

CVC is Git for the AI's brain. Instead of versioning source code (that's Git's job), CVC versions the agent's entire cognitive state — every thought, every decision, every conversation turn — as an immutable, cryptographic Merkle DAG.

Think of it this way:

What developers do with Git	What AI agents do with CVC
`git commit` — save a milestone	`cvc commit` — checkpoint the AI's brain state
`git branch` — try a risky idea in isolation	`cvc branch` — let the AI explore without polluting main context
`git checkout HEAD~5` — go back to an old version	`cvc restore <hash>` — rewind the AI to any previous thought
`git merge` — bring changes back	`cvc merge` — semantically merge AI insights from branches
`git log` — see the history	`cvc log` — see the full history of AI decisions

How It Works Under the Hood

CVC uses a three-tier local storage engine:

SQLite — Stores the commit graph, branch pointers, and metadata. Fast traversal, zero configuration. This is your .git/ equivalent.
CAS Blobs (Content-Addressable Storage) — Every context snapshot is Zstandard-compressed and stored as an immutable, content-addressed blob. Deduplicated and efficient, identified by cryptographic hash.
ChromaDB (Optional) — Semantic embeddings of every commit, enabling natural-language search. "When did the agent solve a similar database migration?" — found instantly.

These three tiers are tied together by a Merkle DAG — the same cryptographic structure Git uses. Immutable. Verifiable. Each commit points to its parent, forming an auditable chain of the AI's reasoning.

Key Features

Save (Cognitive Commits)

Checkpoint the AI's brain at any stable moment. Every conversation message, every decision, every tool call — captured as an immutable snapshot.

bash

cvc commit -m "Auth system design complete"

Branch (Isolated Exploration)

Try risky approaches without polluting the main context. If it works, merge back. If it doesn't, switch back — zero damage done.

bash

cvc branch experimental-refactor
# AI explores the risky approach...
cvc merge experimental-refactor  # bring back the good parts

Rewind (Time Travel)

Stuck in a loop? AI keeps making the same mistake? Rewind to any previous checkpoint — one command.

bash

cvc restore a1b2c3d4  # back to before the loop started

Merge (Semantic Integration)

CVC doesn't just concatenate text. It performs semantic three-way merges — intelligently combining insights from different conversation branches.

Search (Knowledge Recall)

Search across all past conversations by meaning. "When did we solve that CORS issue?" — answered in milliseconds.

bash

cvc recall "CORS configuration fix"

Research-Backed Results

CVC isn't just theory. It's backed by published research:

Metric	Result	Source
58.1% context reduction	Via branching	ContextBranch paper
3.5× success rate improvement	With rollback	GCC paper
~90% cost reduction	Through prompt caching	Anthropic/OpenAI caching
~85% latency reduction	On restores	Anthropic prompt caching

Three Ways to Use CVC

1. Built-in CLI Agent

Just type cvc in your terminal. You get a powerful AI coding assistant with 17 tools, 31 slash commands, and full Time Machine capabilities.

bash

cvc                          # Start the agent
cvc agent --provider openai  # Use GPT instead of Claude
cvc agent --provider ollama  # Use free local models

2. Universal Proxy (for any AI tool)

CVC sits between your favorite AI tool and the LLM provider as transparent middleware. Claude Code, Aider, Cursor — they all get a time machine without knowing it.

bash

cvc launch claude    # Claude Code with time machine
cvc launch aider     # Aider with time machine
cvc launch cursor    # Cursor with time machine

3. MCP Server (for IDEs)

CVC exposes 20 tools via Model Context Protocol for VS Code + Copilot, Cursor, Windsurf, and any MCP-compatible IDE.

bash

cvc mcp              # Start MCP server

How CVC Compares

Feature	Claude Code / Codex	Cursor / Copilot	CVC
Context after restart	Lost	Lost	Restored from checkpoint
Branching	None	None	Full branch support
Rewind to any point	Single undo	Ctrl+Z	Any prior commit
Search AI history	Current session	Session only	Semantic search all sessions
Provider lock-in	Anthropic only	Vendor-specific	5 providers, fully agnostic
Cost	Full re-prompting	Full re-prompting	~90% cheaper via caching
Offline support	No	No	Yes (Ollama + local models)

Get Started in 30 Seconds

bash

# Install
pip install tm-ai

# Or one-line installer
curl -fsSL https://jaimeena.com/cvc/install.sh | bash   # macOS/Linux
irm https://jaimeena.com/cvc/install.ps1 | iex           # Windows

# Launch
cvc

That's it. Two commands. Your AI agent now has a time machine.

What's Next?

CVC is actively developed. Coming soon:

Cloud sync for team collaboration (Q3 2026)
Team shared memory — pooled AI knowledge base
Metrics dashboard — team-wide analytics

Check the full documentation at jaimeena.com/cvc/docs and the installation guide at jaimeena.com/cvc/install.

CVC is free to use and available on PyPI as tm-ai.