Security and local-first notes

Local-first

The alpha workflow stores MemoryGuard data locally. No cloud service and no external LLM API is required for the core capture, memory, and Context Sync path described in these docs.

SQLite under .memoryguard

The local store uses SQLite under .memoryguard. Treat that directory as project-local state and decide deliberately whether it belongs in version control for your workflow.

Best-effort deterministic redaction

MemoryGuard attempts to detect and omit secret-looking content before rendering generated context. This is a guardrail, not a guarantee. Review AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, MEMORY.md, and the Cursor rules before publishing.

Transcript hygiene

Do not paste real secrets into transcripts when avoidable. If a session file contains credentials or sensitive customer information, clean it before capture or reject the sensitive candidate during review.

Trust and provenance

Every memory carries a source, a scope, a status, a sensitivity tier, and a trust score. memoryguard show <id> prints the full breakdown, and the Context Sync engine only renders memories that pass the trust, scope, and policy filters.

Open-core boundary

The OSS core is the part you are looking at. The open-core boundary is enforced in code: memoryguard-core and memoryguard-models never import from a commercial package, and any commercial behavior enters the core only through stable interfaces gated by feature flags that default off.