Time Machine for AI Agents

Cognitive Version Control

AI forgets in 20 minutes. We remember forever.

v3.5.39🐍Python·3.11+
curl -fsSL https://jaimeena.com/cvc/install.sh | bash

Quick Navigation

Your AI coding agent is brilliant...
for about 20 minutes.

Then it forgets what it already fixed, contradicts its own plan, and loops on the same error for eternity. A bigger context window doesn't fix context rot—it just gives it more room to spread.

Permanent Memory for Your AI Agent

Like Git, but for the AI's brain. CVC saves the agent's entire context—every thought, decision, and conversation turn. No more forgetting. No more loops.

Save

Checkpoint the agent's brain at any stable moment. Never lose a good thought process.

Branch

Explore risky ideas in isolation. Keep your main context clean and focused.

Rewind

Stuck in a loop? Time-travel back instantly to a state before the error occurred.

Merge

Merge learnings back—not raw logs. Combine successful experiments into one.

cvc
cvc gateway start

Just type cvc and you're in.

CVC Agent CLI Interface

CVC Self-Improvement Stack — 6 Layers

Built on an immutable Merkle DAG. Every layer below is enabled by default. None of them require user input to start.

1Adaptive Behavior
Injects user-model + predictive-loader context into every system prompt. Learns from every session.
2Dream Cycle → Skill Auto-Generation
After every turn, drafts new skills from the conversation and queues them for review.
3Weekly Patch Proposal Engine
Gated code modifications: propose → sandbox → auto-test → surface diff → user approves via the dashboard.
4Policy Gates
Declarative YAML allow/deny/require_approval rules — drop files into ~/.cvc/policies.d/ for instant guardrails.
5Mode Switch
observe (no writes) / supervised (default) / autopilot (full trust). Enforced at the dispatch layer, not just hinted.
6Mission Envelopes
Long-running agent jobs with budget caps (wall / tokens / USD / iterations) and a fallback chain of 5 LLM providers.
SHIPPED August 2026

What's new in v3.5.39

Every release tightens the loop between the model, the user, and the system. This one fixes three cross-platform bugs that broke the agent's tool calls on Qwen/MiniMax wire-format output.

FIX

Gateway text-salvage

The dashboard chat path now recovers tool calls that the model emitted as wire-format XML () and surfaces them as real tool_start + tool_result SSE events. Wire-format wrappers are stripped from the streamed text.

UX

Silent dud recovery

The yellow "⚠ tool called with no arguments" warning is suppressed entirely. The model gets a nudge to retry with real arguments and the user sees the corrected call on the next turn — no scary noise on Termux, Mac, Windows, or Linux.

UX

One-question-per-clarification

The ask_user tool description now explicitly forbids pre-generating multiple questions. The model asks ONE question, waits for your answer, then asks the next.

DIAG

cvc doctor-env

New command: tests cogs, deps, config, AND a live LLM ping in one shot. Reports ✅/⚠️/❌ with one-line fix hints. Exits non-zero on any ❌ so it works in CI.

FIX

Anthropic /api/chat base_url

The dashboard's MiniMax / Anthropic-compatible route now resolves to https://api.minimax.io/anthropic. Six regression tests pin the URL across five resolution layers.

DOC

Cross-platform install

One-line installer now works identically on macOS (Apple Silicon + Intel), Windows 10/11 + WSL2, Linux x86_64, Linux ARM64 (Pi 4/5, Graviton), Termux on Android, and Docker (linux/amd64 + linux/arm64).

Explore CVC

Deep-dive into every aspect of CVC — from quickstart guides to enterprise features.