Everything a team needs.
Nothing it doesn't.

Recall that just works
πŸ›οΈ

Shared workspace memory

A first-class wings registry organizes drawers into workspaces β€” backend, frontend, data, infra. Wings auto-register the moment an agent writes a memory, so every workspace inherits its context automatically.

wings Β· rooms Β· drawers
🧭

Project-aware agents

Agents detect whether the current repo is a known project or topic and act on it β€” palace_project_status reports the state, palace_create_wing declares a topic wing, and palace_list_wings returns the full registry.

project_status Β· create_wing
πŸ”

Hybrid semantic search

The same BM25 + cosine retrieval from palace-rs, now over your whole team's collective memory, with source-grounded provenance on every hit. R@5 = 0.981 on LongMemEval.

recall@5 = 0.981
Shared across your team
πŸ•ΈοΈ

Shared knowledge graph

Temporal entity relationships across all teammates. Facts written by one agent are queryable by all β€” architecture decisions, API contracts, team conventions.

KG Β· temporal Β· facts
πŸ“–

Team diaries

Session diary entries in AAAK format β€” private by default, optionally shared. Any agent can warm-start from the last person's session context.

AAAK Β· warm-start
πŸ”

Workspace permissions

Role-based access (owner, admin, maintainer, developer, read-only agent, security auditor) with each organization's data isolated by Postgres row-level security.

RBAC Β· RLS Β· seats
πŸ“Š

Usage tracking + audit

Every MCP tool call is recorded as a usage event, and every memory write lands in a tenant-scoped audit log. Gain reports and Prometheus metrics show adoption across the team.

palace_gain Β· /metrics
Runs anywhere. Trusts no one.
πŸ“‘

MCP server β€” drop-in

Exposes all palace_* MCP tools over the network. Any MCP-compatible agent β€” Cursor, Claude Code, Codex β€” connects with zero configuration change.

MCP Β· stdio Β· http
πŸ¦€

Built in Rust

Single binary, low memory, fast startup. Runs comfortably on a small VM or a Raspberry Pi. No JVM, no Node, no Python runtime needed.

~10 MB binary
πŸ”’

Restricted-network friendly

Embeddings run locally via ONNX. License verification works fully offline. The Docker image includes the embedding model, so it runs with no internet access at all.

no cloud dependency

Give your agent a
memory today

Free and local in two minutes. Shared across your whole team in ten. No cloud, no lock-in.