claude-mem has an LLM summarize your sessions into observations. Palace stores what was written, verbatim, searches it locally, and shares it across a team on your own infrastructure.
Local ONNX embeddings: no third-party embedding API required.
Ed25519-signed licenses verify offline for air-gapped operation.
Palace vs claude-mem, side by side
Feature
Palace
claude-mem
How memories are written
Verbatim text, stored as written. No LLM in the loop
An LLM observer writes observations and summaries
Model provider
None. Embeddings run locally on ONNX Runtime
A hosted observer, an OpenRouter or Gemini key, or your Anthropic plan
What you run
One binary and one SQLite file; the team server adds Postgres
Node, Bun, uv (Python), a background worker, and Chroma by default
Telemetry and sign-in
No telemetry and no account for the local engine
Telemetry on by default (opt-out); the installer asks for an email sign-in
Published retrieval result
98.7% recall@5 on LongMemEval, with the method and command
No retrieval benchmark in its README
Shared team memory
Palace Server: self-hosted, with roles, tenant isolation, and a write audit
Check its current documentation
Two ways to remember
claude-mem runs an observer model over your sessions and keeps what it writes. Palace keeps the original text of decisions, fixes, and conversations, so the detail an agent needs months later is still there, and every answer cites the drawer it came from.
What sits in the loop
With claude-mem, a model provider processes your session content to produce memories. With Palace, embeddings are computed on your own CPU and nothing is summarized by an LLM, so no session content goes to a provider to become memory.
Where Palace fits
Choose Palace when memory has to stay inside your network, be shared across a team with access controls, and be backed by a retrieval result you can reproduce. Details on this page reflect each project's public documentation in November 2026.