Known limits

Breno is early, and a document that only listed what works would be a poor guide. This page lists what does not work yet, or has not been proven, grouped by area. It is kept current with the project's own plan of record.

Scale#

• A thousand connected sources is a goal, not a result. A full run with 1,000 connections completed every sync and lost no jobs, but failed its latency budget: 95th-percentile schedule lag of 327.66 seconds against a budget of 300, and 500 jobs that ran past the two-minute job timeout. The run did not route through a real Nango proxy, and the machine suspended for about five hours during it. A controlled re-run is still to be done.

Sources and sync#

• Seven apps bring knowledge in (Slack, Google Drive, Gmail, Notion, GitHub, Jira, Linear). Others connect only. Confluence, SharePoint and OneDrive, Dropbox and HubSpot do not ingest yet.
• The priority sources have no live end-to-end tests against real accounts; they are tested with fixtures and fakes.
• Some providers give no deletion signal (Teams, Linear, GitLab, Asana, ClickUp, HubSpot, Salesforce). In Slack and Teams, messages arrive as per-sync digests.
• Full listings are capped per provider; beyond the cap, deletions are not propagated.
• Reconnecting a native OAuth connector picks the first connector of that provider when there are several.
• Bulk import takes Markdown and text zips only, up to 200 files.
• Links that need JavaScript to render are not supported. Scanned PDFs are read from embedded page images, up to 30 pages.

Access#

• Source permissions are not read. Connector items arrive workspace-visible with sensitivity internal.
• A shared entity's description can still be drawn from an item you cannot read, and the graph's search limit applies before filtering.

Retrieval and answers#

• On the test set, the right document is in the top three for 83.3% of questions. The faithfulness measure of generated answers, with a target of 90%, is not implemented.
• Reciprocal-rank fusion measures better and is deliberately not used, because it weakens the refusal guard.

Memory#

• You cannot record a decision from the interface, and decisions are not extracted automatically from architecture decision records. They come from meetings.
• Memory records and decisions are still a separate store from items.
• After an item is processed again, the graph store can keep old evidence links until the item is purged.

MCP and API#

• The OAuth flow for Claude.ai and ChatGPT is built but has not been exercised against the live services; the MCP Inspector has not been run; cursor:// redirects are not accepted.
• The TypeScript SDK does not reconnect MCP sessions by itself (the CLI bridge does).

Audit and retention#

• The audit log covers organization and security events, exports and imports, and meetings. Item create, delete, share and space changes, and searches, are not yet logged.
• Retention and legal hold apply to search history, agent tasks and routine runs, not yet to knowledge items.

Backups#

• Scheduled workspace backups are not built. You can export on demand, and operators back up with their own tooling. See Backups and restore.

Mobile#

• The mobile apps capture text, links, voice and photos into Knowledge, but have no "Only me" option, no item detail screen and no share extension.

Zoom#

• The Zoom recording hook works in tests with simulated webhooks; it has not been exercised against a live Zoom account.

Certifications#

• Breno holds no certification (such as SOC 2 or ISO 27001). A control mapping exists as evidence for an audit, and is not an audit. See Compliance.

Found something not listed here?#

Tell us at contact@breno.ng.