Every read and write operation in Palimp is scoped to a namespace. There is no mechanism to query across namespaces. Inserting data into namespace A and querying namespace B will return zero results.

When Palimp exposes memories and knowledge through MCP tools, the output is always structured data. It is never presented as an instruction to the consuming model. The safety.treat_as_instruction field is always false.

Palimp stores all data in a local SQLite database. No data is sent to external services unless an optional HTTP embedding provider is explicitly configured. The default configuration is fully offline.

Deleted sources are tombstoned, not physically removed. Tombstoned episodes and their dependent graph facts are excluded from all recall queries. Audit logs are preserved.

Palimp is designed for single-user or small-team local use. It does not provide:

• Multi-tenant authentication
• Network-level encryption
• Role-based access control
• Rate limiting beyond what the OS provides

For production multi-user deployments, place Palimp behind a reverse proxy with authentication.