Limitations and future work

OrgMemory is a production-shaped POC, not an approved enterprise deployment. The following constraints are intentionally explicit so architecture and test coverage are not mistaken for operational certification.

Current limitations

Operational evidence

• database and object-storage restore rehearsal remains open;
• malware and DLP scanning are not integrated into the upload path;
• production load, latency, capacity, and high-concurrency revocation timing have not been measured against service-level targets;
• provider-backed retrieval quality evaluation and an enterprise security review remain open;
• complete audit/export leak coverage across every retrieval surface is not yet claimed.

Connectors and identity

• Slack, Google Drive, and GitHub adapters are extensively tested with recorded responses, but current evidence does not include a live production workspace proof;
• incremental webhooks, credential-rotation automation, OCR, Airbyte staging, and malware/DLP integrations are not implemented;
• SCIM discovery and the protected connection foundation exist, but User and Group mutation endpoints remain disabled;
• one external issuer/subject belongs to one application user and organization; one person across multiple organizations needs a future global-person model;
• source-derived permissions for content composed from multiple source objects do not yet have a complete intersection contract.

Product and API

• general Assistant chat-turn idempotency is not implemented;
• historical conversation answers remain snapshots; citation opening rechecks current access, while purge-on-revocation would require a separate retention policy;
• administration screens and the graph explorer lack focused real-browser coverage beyond their API/service and shared-shell tests;
• controlled SOP effectivity, Skill package replacement/removal, public package signing, and a cross-company marketplace are deferred;
• public product versioning and a compatibility policy are not established.

Retrieval and evaluation

• multi-space candidate-level reranking remains fail-closed until one global contract exists;
• graph strategy selection is internal rather than a public tuning surface;
• the optional VChordRQ index requires an extension absent from the pinned local image and fails startup when selected;
• one preserved public evaluation fixture contains a known expected-denial inconsistency; the product does not hard-code an exception for it.

Evidence-gated future work