The most consequential thing shipped this week for anyone building internal knowledge systems isn't a model or a retriever. It's an ingestion path that routes around both.
On August 13, OpenAI added Computer History to ChatGPT Business for the macOS app, letting members bring context from selected apps and websites into ChatGPT and Codex — an opt-in feature that turns activity across apps and websites into memories and a timeline the assistant can use . Users choose which apps and websites contribute, can pause collection, and can review or delete history at any time; the feature is available to Pro, Business, and Enterprise, with Business and Enterprise admins required to enable access first, and initial availability excludes the EEA, Switzerland, and the UK.
Why this breaks the connector assumption
Every enterprise knowledge stack built in the last three years rests on the same premise: content lives in systems of record, you index it through connectors, and you inherit the source system's permissions at query time. Document-level ACLs are the whole reason permission-aware retrieval works.
Endpoint capture inverts that. The unit of ingestion is no longer a document with an owner and a sharing policy — it's whatever rendered on one employee's screen. A Salesforce record a user shouldn't have exported, a competitor's pricing page, a Figma comment thread, a half-finished doc from another team's drive: all of it becomes retrievable context with no provenance chain back to the originating ACL. Your careful permission mirroring doesn't apply to an index you don't operate.
The lifecycle problem is just as real. Retention in a corporate knowledge base is a policy decision made centrally. Here, deletion is a per-user control. That's good privacy design and bad records management, and the two are genuinely in tension.
Worth studying, though, is the scoping model. Shared projects use project-only memory and cannot be switched to default memory, and ChatGPT Work is not available in projects using project-only memory. That's a hard namespace boundary rather than a relevance-ranking heuristic — the right primitive if you're designing memory isolation, and one most agent-memory libraries still fudge with metadata filters.
Practical read
If you run internal search, plan for a shadow index you can't audit. Recall stops being your differentiator; provenance, permission enforcement, and query-time auditability become the things you can offer that endpoint capture structurally cannot.
Separately, the MCP stateless migration has entered its debugging phase. Anthropic's recent notes include a fix for MCP v2 connections endlessly reopening the subscriptions/listen stream against servers that terminate long-held streams on a fixed timeout, such as serverless hosts . That's a predictable consequence of the 2026-07-28 spec — the stateless core means MCP servers can run in a Worker with no stateful infrastructure, with Tasks providing the path for reliable long-running work . If your server sits behind a serverless host, check your stream timeouts against client reconnect behavior now.