# MEMORY.md - Miegyver

## Deployment facts (reviewed 2026-11-03)

- Active team: Miegyver platform/operations and Thinker QA; Theo is the human operator.
- Kubernetes deployment, not legacy Compose/autoheal/Ollama procedures.
- R2 holds agent workspaces, shared local-kb, bridge YAML and a non-secret OpenClaw
  bootstrap. The initializer copies bootstrap only when active JSON is absent.
  Active writable JSON, sessions and SQLite remain on the filesystem state PVC.
  R2/bootstrap edits do not automatically replace existing active JSON.
- L4/pgvector memory and embeddings are not implemented. Do not infer retrieval tools
  from historical memory architecture descriptions.
- Role-relevant MCP reads are enabled: Grafana dashboards/Prometheus metrics,
  GitHub relevant source/PR/CI status and platform-kb search/read. Platform-kb
  create/update is enabled for task-relevant sanitized reports and knowledge;
  archive/delete, native writes, raw logs/traces and secret-bearing content remain denied.
- Caller-ID filtering was explicitly deferred. Channel access is not authorization
  for sensitive data. Bridge masking is deployment-dependent; verify, never assume.

## Notes

This file holds verified durable context, not procedures or an alternative policy.
Local memory is operator-curated; KB MCP publishing does not grant local writes. Procedures
and references live in local-kb/sop/. Useful historical evidence remains in the
operator backup, not current operating instructions.

- Lessons: runbooks describe procedures; ADRs record reviewed architecture decisions. Validate config against the installed runtime schema.