Vant Documentation
Agent memory that lives in your repo.
Featured: Evolution Without Drift - the agent-first white paper. Memory as a portable spec, continuity across generations of agents, and the methodology that built Vant - written by the agents, with every claim linked to a passing test.
Vant gives AI agents a brain: memory files the agent reads when it wakes and writes when it sleeps. The files are plain markdown in your own repository. Git history becomes the memory lineage. Any agent on any stack can use it, because the memory is files and the tool interface is MCP.
Think of it as your agent’s soul that reincarnates with full memories. What
that means in practice: on wake, the agent loads identity, goals, lessons,
and errors from models/; on sleep, it writes what it learned back. The
next session starts with all of it.
Three memory systems
| System | What it stores | Where |
|---|---|---|
| The brain | Markdown memory files: identity, goals, lessons, errors, preferences | models/public/<brain>/ and models/private/<brain>/ |
| The memory store | Key-value state and documents, with TTL and sandbox gating | storage layer, per-brain scoped |
| Brain search | Semantic retrieval over the corpus: embeddings, rerank, citations for grounding | embeddings over brain files |
Everything is in your repo. Nothing is in someone else’s database. If you
stop using Vant, the memory is still there, readable with cat.
Quick paths
For humans: install and start
Vant is not on npm (the vant package there is an unrelated Vue UI
library). Clone and install:
git clone https://github.com/dhaupin/vant.git && cd vant
npm ci
Start the full sequence: layout check, health, ready:
node bin/vant.js start
vant start also imports an old single-brain layout automatically if it
detects one. See the migration guide.
For agents: connect over MCP
Start the MCP server with auto-wired tools:
vant mcp
The server listens on 127.0.0.1:3457 by default. Point any MCP client at
it and the brain surface arrives as tools: read, write, search, migrate
status. Full client setup in MCP.
Agents inheriting a repo with an existing brain need one command before anything else:
vant migrate --status
It reports the brain layout version and whether a legacy import is pending.
Why files in git
- Ownership. Your repo, your history, your backup story.
- Review. What your agent learned shows up in diffs. Approve lessons in a PR, revert the bad ones.
- Branching. Run an agent crew with one branch per agent, then merge what is worth keeping.
- Portability. No API, no export step, no lock-in. The memory is files.
Runtime underneath
Memory is the product; the runtime makes it safe to use unattended. A security chain gates every brain operation (sandbox capabilities, input validation, rate limiting, escrow approval). A write-ahead journal and atomic writes survive crashes. Metrics and health endpoints expose what the system is doing. Start at Runtime.
Multi-brain and multi-agent
One Vant install holds several named brains with a stack you switch between. Agents get their own branches and work in isolation. Trust levels and a succession ledger control how much state each generation inherits. See Multi-brain and Succession.
Sections
| Section | Contents |
|---|---|
| Getting started | Install, configure, first run, agent onboarding |
| Memory | The brain, memory store, search, citations, horcrux, geometry |
| Runtime | Programmatic API, MCP, headless server |
| Multi-agent | Brains, branches, succession, crews, federation |
| Operations | Storage, journal, events, cache, CI |
| Security | Sandbox, gates, escrow, sudo |
| Integrations | GitHub, agent skills, Linear, Docker, S3 |
| Reference | CLI commands, configuration, changelog |
| Advanced | Search internals, NSC9 spec, style, release process |