FAQ

Is Codifix a vector database?

No. It uses PostgreSQL and can use pgvector, but the product is a memory workflow: capture, curation, evidence, retrieval, briefings, agent integration, and evaluation.

Does it replace documentation?

No. It complements docs by preserving decisions, incidents, project conventions, and operational context that often never reaches README files.

Does it work offline?

The service and deterministic retrieval paths can be local-first. External LLM or embedding providers receive text only when you configure and use them.

Can I use local embeddings?

Codifix supports OpenAI-compatible embedding endpoints, which can include local compatible servers when configured.

Can multiple projects share one database?

Yes. Queries and memories are project-scoped. Use clear project slugs and inspect status when moving repositories.

An operator-configured area is a separate service/data boundary, not another project slug. Global search searches accessible projects in the selected service; it does not automatically query other areas or private federation collections.

Is MCP entirely read-only?

Ordinary memory tools are read-only. Current source also exposes loop control-plane actions and dedicated expertise submission tools. Private expertise submission is an external action with separate consent/service policy and model-output gates. See MCP.

Does restarting upgrade a named instance?

No. A named instance can pin its own release. Use its reviewed platform upgrade procedure; package installation and service restart-all do not replace that pin. The current source version 3.0.0 alone does not identify every source increment present in an installed artifact.

What should become a memory?

Durable project facts, decisions, conventions, fixes, operational lessons, and constraints. Do not memorize transient observations like "I opened a file".

How do I know it works?

Run status and ask a question whose answer should require project context. Inspect citations before trusting the answer.