Xavier — Fast Vector Memory for AI Agents

Xavier is a Rust-based memory runtime for AI agents with HTTP, CLI, and MCP entry points. It stores, retrieves, and manages vector embeddings, hierarchical memory, and semantic relationships over a SQLite-backed store, giving agents fast contextual recall without external dependencies.

Quick Start

# Option 1: Install from source
cargo install xavier

# Option 2: Run with Docker
docker run -p 8006:8006 -v xavier-data:/data ghcr.io/iberi22/xavier:latest

# Start the server
export XAVIER_TOKEN=your-secret-token
xavier http

# Add and search memory
xavier add "AI agents should always verify their sources" "agent-guidelines"
xavier search "agent guidelines"

Installer

Xavier ships with an interactive TUI setup wizard that configures everything in 6 steps — no manual config editing needed.

One-liner install

Windows (PowerShell):

irm https://raw.githubusercontent.com/iberi22/xavier/main/install.ps1 | iex

Linux/macOS:

curl -fsSL https://raw.githubusercontent.com/iberi22/xavier/main/install.sh | bash

Key Features

• Belief Graph & GraphRAG — Hierarchical memory with semantic relationship mapping for deep context.
• Native Security Scanner — Built-in multi-layer protection against prompt injection and sensitive data leaks.
• HTTP API — JSON REST endpoints for memory CRUD with token-based auth.
• CLI Client — High-performance add, search, and stats commands.
• MCP Server — stdio-based Model Context Protocol server.
• TUI Setup Wizard — Interactive installer (6 steps) for Windows, Linux, and macOS.
• Service & Daemon Support — systemd (Linux) and Scheduled Tasks (Windows) for background execution.
• Chronicle Workflow — Automated harvesting of project activity to generate technical documentation.
• Plugin System — Extensible enterprise integrations (PgHeart).

┌─────────────┐  ┌──────────┐  ┌──────────┐
│   CLI       │  │  HTTP    │  │   MCP    │
│  (add/search)│  │  Server  │  │  (stdio) │
└──────┬──────┘  └────┬─────┘  └────┬─────┘
       │              │              │
       └──────────────┼──────────────┘
                      │
              ┌───────▼────────┐
              │  Core Engine   │
              │  (add, search, │
              │   stats,       │
              │   export)      │
              └───────┬────────┘
                      │
              ┌───────▼────────┐
              │  SQLite Store  │
              │  + Vector      │
              │  Embeddings    │
              └────────────────┘

The three entry points (CLI, HTTP, MCP) share the same core engine, which handles memory operations over a SQLite-backed store. Each entry point is independent — you can run the HTTP server, use the CLI against it, or connect the MCP server to any MCP-compatible host.

Public Dataset Export

Generate a public, read-optimized dataset for agent context without cloning or rebuilding:

xavier export --public

Output lives in xavier-dataset/ at the repository root with NDJSON files for memories, entities, timeline events, git commits, code symbols, and more.

Example agent bootstrap from GitHub raw:

BASE="https://raw.githubusercontent.com/iberi22/xavier/main/xavier-dataset"

curl -fsSL "$BASE/dataset_manifest.json"
curl -fsSL "$BASE/memories.ndjson" | head -n 20
curl -fsSL "$BASE/code_symbols.ndjson" | jq -c 'select(.kind == "function")' | head

Full export schema is documented at docs/FEATURE_STATUS.md.

HTTP API

curl http://localhost:8006/health

curl -X POST http://localhost:8006/memory/add \
  -H "X-Xavier-Token: $XAVIER_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"content":"Design decision: use RRF","path":"decisions/001"}'

curl -X POST http://localhost:8006/memory/search \
  -H "X-Xavier-Token: $XAVIER_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"query":"design decision","limit":5}'

Full API reference: docs/site/src/content/docs/reference/api.md.

MCP

Start the MCP stdio server:

xavier mcp

Current MCP tools: search, add, stats.

Public Data & Export

Xavier will expose a public export pipeline through:

xavier export --public \
  --huggingface-repo iberi22/xavier-dataset \
  --huggingface-token $HUGGINGFACE_TOKEN

The export protocol splits lightweight public context from heavy analytical artifacts:

1. Generate NDJSON manifests and JSON schemas, then commit them to GitHub in iberi22/xavier-dataset.
2. Generate Parquet files for embeddings and metrics, a complete .sqlite3 snapshot, and vector indexes such as .lance/ or .faiss.
3. Upload the heavy artifacts to the Hugging Face dataset iberi22/xavier-dataset.
4. Include Hugging Face artifact URLs inside the NDJSON records committed to GitHub.

Use GitHub raw URLs for lightweight agent context and Hugging Face for larger downloads.

Layer	Location	Contents	Typical size
Manifest + context	GitHub raw	NDJSONs, schemas	~1-10 MB
Analytical data	Hugging Face	Parquet files for embeddings and metrics	~50-500 MB
Database + vectors	Hugging Face	.sqlite3, .lance/, .faiss	~100 MB-2 GB

See the full public export reference in docs/site/src/content/docs/reference/export.md.

Configuration

Runtime configuration lives in config/xavier.config.json. Secrets go in .env (see .env.example).

Variable	Default	Description
XAVIER_TOKEN	required	Authentication token for HTTP API
XAVIER_CONFIG_PATH	auto-resolved	Path to xavier.toml. Defaults to XDG_CONFIG_HOME/xavier/xavier.toml or config/xavier.config.json.
Provider keys	unset	External API credentials (e.g. embedding providers)

Documentation

• Agent Rules — Canonical instructions for optimal agent memory integration
• Installer & Setup — TUI wizard screenshots and install guide
• Feature Status — Current verified surface and 1.0 gaps
• CLI Reference — Full command documentation
• API Reference — HTTP endpoint details
• Architecture — System design and hexagonal architecture
• Public Release Roadmap — Upcoming milestones

Status

Current release: 0.6.1-beta (Development). All core systems (Security, Belief Graph, Hierarchical Memory) are stabilized and verified.

License

MIT — see LICENSE for details.