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.

Entropy Patching 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 68 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.

Islands are lazy-loadable brain components (v0.8.6+). Instead of loading all 68 brain files on startup, you load only what you need: github, gitlab, linear, automation, stego, resolution. This dramatically reduces cold start time and memory footprint. Islands are defined in models/public/identity.md and loaded via vant islands --load .

Stego is Horcrux/steganography bootstrap (v0.8.6+). Embed your brain data in images or other media. This provides true deniable persistence - your AI's memory exists but is invisible to casual inspection. Useful for AI agents operating in adversarial environments where plain-text memory is a liability.

Adaptive Entropy 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.

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.

No - Vant works with any Git provider: GitHub, GitLab, Bitbucket, or self-hosted. Use what you already have. A free GitHub account works perfectly.

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.

Three layers: cache (same query = instant), compact (summaries skip rehydration), lazy-load (heavy modules load on-demand). Scales to 10k+ memories.

RAG rerank (v0.8.6+) keyword-scores memories against your query and compresses them to fit token budgets. It's different from search: search finds semantic matches, rerank prepares them for LLM context. Use rerank to optimize memories before sending to your AI - it strips markdown fluff and truncates to your token budget.

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.

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.