Shared memory for AI coding agents, self-hosted on your infrastructure. Describes Palace Server v0.6.0.
Palace Server gives a team's AI coding tools (Cursor, Claude Code, Codex, Claude Desktop) a shared, persistent memory over MCP. It is a single Rust service (Docker image or binary) backed by your own PostgreSQL with pgvector. Nothing is hosted by us.
| Direction | To | When | Data sent |
|---|---|---|---|
| Inbound | Palace Server, port 8080 (HTTP) | Developers' AI tools (MCP) and the admin UI | Memory reads and writes, with a bearer API token |
| Outbound | huggingface.co | Never with the Docker image, which includes the embedding model. Binary installs fetch it once unless pointed at a local copy | None of your data; a plain file download |
| Outbound | api.palacememory.com | Off by default. If enabled, once a day to renew the license | The license key only |
| Outbound | Your OIDC identity provider | Only if SSO is configured, during sign-in | Standard OIDC sign-in requests |
Nothing else: no telemetry, no analytics, no update checks. Logs go to stdout. Serve port 8080 behind your reverse proxy, which terminates TLS.
pg_dump.sslmode=verify-full). Encryption at rest is provided by your database and disk setup.ghcr.io/ancientice/palace-server (linux/amd64, linux/arm64) with the embedding model included, so it runs with no internet access. Binaries for Linux (x86_64, arm64) and macOS (arm64)./healthz, /readyz.Free 60-day evaluation license for up to 10 seats, no credit card. Start it at palacememory.com/pricing. For a larger pilot or security questions, write to help@palacememory.com. Install guide: palacememory.com/docs.