Persistent AI Memory
Your AI should remember. Vant is an open source system for AI agent memory persistence via GitHub. Each session inherits the full context from previous ones.
v0.8.4Brain files live in Git. Version control built in. Pull changes from previous sessions automatically.
Learn more →Token-aware compression. Separates high-entropy data from stable patterns. Dramatically reduces LLM prompt size.
Learn more →Self-calibrating threshold (μ + k×σ). Auto-detects optimal compression for your data.
Learn more →Stable patches include anchors. Tells the LLM where to place high-entropy data in context.
Learn more →Track thoughts as resolved, deprecated, or rejected. Frontmatter updates directly in brain files.
Learn more →Branches and locks prevent conflicts. Multiple agents can work on the same brain without overwriting each other.
Learn more →Official Docker image builds on release tags. Run anywhere. Portable brain survives container restarts.
Learn more →Memory tools via JSON-RPC. AI agents read/write brain over HTTP. Standard MCP protocol.
Learn more →Experiment with different brain configs. Merge via pull request. Always keep a working main branch.
Learn more →Vant v0.8.4 solves a fundamental problem: AI agents lose all context when sessions end. Traditional memory systems require complex databases or external services. Vant uses Git as a simple, reliable foundation.
Why Git? It already handles versioning, branches, conflict resolution, and distributed sync. Every commit is a checkpoint. Every branch is an experiment. Every pull request is a review.
With Entropy-Patch (v0.8.4+), Vant evolves from "Context Storage" to "Latent Transport"-token-aware compression that makes long-running autonomous agents practical.
Use Vant for:
The brain is your AI's memory and personality. Vant makes it persistent, portable, and versioned.
Full documentation →Your AI's memory-a folder of markdown files that store who you are, what you've learned, goals, preferences, and context. Think of it as your AI's personality and knowledge base. Each session loads the brain from GitHub, makes changes, and pushes back. The next session inherits everything automatically. It's like human memory but versioned.
Learn more →Entropy-Patch is Vant's token-aware compression (v0.8.4). It scans your brain files and separates "stable" repetitive content from "spiky" high-entropy data. The result? Massive token savings. Instead of sending 19 markdown files to your LLM, you send one compressed .vpatch file. Includes Semantic Seeds so the AI knows where each piece of data belongs in context.
Learn more →AdaptiveEntropy automatically calibrates your compression threshold using statistical analysis (μ + k×σ). Instead of guessing the right threshold, it analyzes your data and picks optimal settings. Lower k = more aggressive compression, Higher k = preserve more. Great for long-running autonomous agents that need to "remember" correctly without hitting token limits.
Learn more →Each time your agent updates the brain, Vant tracks entropy drift between generations. Low delta = stable memory (good). High delta = context shift (warning). This helps you spot when your agent is "forgetting" or when its personality is drifting. Built for multi-session autonomy where the agent needs to evolve without losing its core identity.
Learn more →Yes, GitHub is required. It's Vant's storage, version control, and sync mechanism-all in one. Every commit is a checkpoint. Every branch is an experiment. Every pull request is a review. Using a private repo keeps your brain private. A free GitHub account works perfectly.
Learn more →Yes-if you use a private GitHub repository, your brain stays completely private. Vant doesn't use any external databases or services. Your brain lives in your GitHub repo, and only you (or agents you authorize) can access it.
Yes! Vant is built for multi-agent coordination. Use branches for isolation and file locks to prevent conflicts. Each agent works on their own branch, commits changes, and merges via pull request. Safe concurrent agentic workflows without overwrite fears.
Learn more →Vector databases store embeddings (mathematical summaries of text). Vant stores full context-the actual markdown, personality, goals, lessons learned. This means complete session inheritance, not just semantic search. Your agent literally inherits its previous thoughts, not just "similar" ones.
No-Vant is completely open source and free. You only need a free GitHub account and your own LLM API keys (OpenAI, Anthropic, etc.). No database fees, no hosting costs, no middleman.