Built on palace-rs

An open foundation.
An inspectable memory.

The MIT-licensed local engine powers retrieval. Palace Server adds the shared team layer and PostgreSQL-backed deployment.

A concrete retrieval result

Measured recall.
With the context attached.

On the LongMemEval s_cleaned evaluation, the exchange-pair configuration retrieves a relevant session within the first five results for 98.7% of evaluated questions. No LLM and no GPU are involved at any stage.

98.7%RECALL @ 5
470QUESTIONS
MITCORE LICENSE
What this number meanspalace-rs engine benchmark; exchange pairs with the default minilm-l6-v2-q8 model; recall_any at session granularity. The 30 abstention questions are excluded, as is standard. It measures retrieval by the engine inside Palace Server.
ConfigurationRecall @ 1Recall @ 5Recall @ 10
Exchange pairs, as Palace stores conversations91.5%98.7%99.6%
One drawer per session89.8%98.1%99.1%

Source: the palace-rs README. Review the method and reproduce it ↗

Local for an individual.
Shared for a team.

palace-rs runs locally with SQLite and provides the memory tools, free for any solo developer. Palace Server is the commercial, self-hosted team service: shared memory, organization access controls, audit records, and PostgreSQL storage.

See team plans

Put your knowledge to work

Give your next session
a head start.

Start your team trial

60 days · 10 seats · No credit card