Not the same project
Palace is not affiliated with MemPalace. Both organize memory as wings, rooms, and drawers, and both store text verbatim. Palace is its own codebase, written in Rust, with its own retrieval, installers, hooks, and team server.
Similar names, separate projects. Palace is an independent Rust memory engine and team server; MemPalace is a Python project built on ChromaDB. Here is how they differ.
| Feature | Palace | MemPalace |
|---|---|---|
| Implementation | Rust. One binary, no runtime to install | Python package |
| Storage | One SQLite file locally; Postgres with pgvector on Palace Server | ChromaDB by default |
| Retrieval | Hybrid search: BM25 keywords fused with local embeddings | Vector search, with an optional hybrid mode |
| LongMemEval recall@5 | 98.7%, with nothing tuned on the evaluated split | 96.6% raw; 98.4% hybrid on its held-out questions |
| Agent setup | palace install wires up 11 agents, with recall hooks and the rule file | Check its current documentation |
| Shared team memory | Palace Server: self-hosted, with roles, tenant isolation, and a write audit | Local only, per its README |
Palace is not affiliated with MemPalace. Both organize memory as wings, rooms, and drawers, and both store text verbatim. Palace is its own codebase, written in Rust, with its own retrieval, installers, hooks, and team server.
MemPalace installs as a Python package and uses ChromaDB as its store. Palace is one binary and one SQLite file: nothing else to start, and a backup is a file copy. Palace Server adds a shared deployment on Postgres with pgvector.
Both projects publish LongMemEval recall@5 without an LLM in the retrieval path. Palace reports 98.7% on the standard 470 evaluated questions, with nothing tuned on that split. MemPalace's README reports 96.6% in raw mode and 98.4% for hybrid mode on its held-out questions. Each project ran its own harness, so rerun both before relying on the gap. Figures as published in November 2026.
60 days · 10 seats · No credit card