How the question came up
The trigger was ordinary: a playbook was meant to describe Andrej Karpathy’s “LLM Wiki” approach — a locally maintained knowledge archive of raw sources and wiki pages, curated by a language model and governed by a fixed schema file (CLAUDE.md). That led to a more important, simple question: where was the document describing this very concept?
The answer was uncomfortable. The playbook was not where its subject suggested it belonged — in a local, version-controlled folder readable by Obsidian — but in Anthropic’s own cloud, as a database row behind an Artifact ID. A document about decentralised, self-directed knowledge management was itself centrally managed. That is not just an irony; it points to a problem anyone handling sensitive client information should consider before turning an AI system into a personal or business memory: knowledge can have at least two homes, each with different rules for access, retention, control and potential liability.
This article compares a cloud-document architecture (Claude Docs, Claude Memory and Artifacts) with a locally maintained second-brain system inspired by Karpathy’s pattern. Where possible, statements draw on Anthropic’s publicly available policies. Where those sources do not support a firm answer, that limit is stated rather than glossed over.
Two different architectures
The difference is not simply “more cloud” or “less cloud”; it is a categorical one. A Claude Doc or Memory entry exists as a database record in Anthropic’s infrastructure. A Markdown file in your own vault exists independently of any AI provider; Claude is a tool there, not the container for the knowledge.
Where does each kind of knowledge live?
“Cloud versus local” is too broad once different kinds of knowledge are separated. They arise in different places and have different persistence rules:
| Knowledge type | Where it originates | Where it persists | Persistence |
|---|---|---|---|
| Raw data (email, transcripts, documents) | A local source or connector | A local RAW folder, if the system is configured that way | Until the user deletes it |
| Derived knowledge as a Claude Doc or Artifact | Synthesised during a session | Anthropic’s cloud once published | Under the applicable Anthropic rules; see retention section |
| Personal or business facts in Claude Memory | From chats or Cowork sessions | Anthropic’s cloud, associated with an account | Until the user deletes it or turns Memory off |
| Automation outputs (briefings, synthesis pages) | Scheduled tasks connecting sources and the vault | Locally, if the routine writes only to RAW/Wiki | Depends on the routine’s design; not guaranteed by the platform |
Retention periods according to Anthropic
According to Anthropic’s Privacy Center, opting in to model training changes the retention period for chats and coding sessions from 30 days to five years, in de-identified form. That is more than 60 times the shorter period. Violations of the Usage Policy can have separate, longer retention periods: two years for inputs and outputs, and seven years for automated classification scores.
These statements concern chats or sessions. They do not provide an equally precise public retention period for every Artifact or Memory entry. The current policy and terms for the specific account and service are what matter.
Who can see what?
“Who has access?” is not a yes-or-no question. It depends on the role and on whether information sits in a cloud resource or a self-managed vault:
| Role | Cloud Docs / Memory | Local vault |
|---|---|---|
| The user | Access through their account and product features | Access through the filesystem or Obsidian |
| Anthropic during processing | Content is processed there | Content may be read or written during an active session |
| Anthropic after processing | Retention under the applicable rules | No persistent copy is intended unless content is uploaded elsewhere |
| Organisation admin on a Team or Enterprise plan | Access may depend on contract and admin settings; not verified here | No access unless the vault is on managed or shared infrastructure |
| Other people (colleagues, clients, public) | Through sharing and account settings; not assessed further here | Through file, device or Git repository sharing |
The only row with a material knowledge gap is organisational administration. Whether and how far admins on Team or Enterprise plans can inspect user content depends on the relevant terms and settings. Anyone who needs a reliable answer should check their contract and admin console rather than treating this article’s assumptions as evidence.
The concrete case: six automation routines
The second-brain concept under discussion includes six planned routines: a daily briefing from email, calendar and SharePoint; evening meeting preparation; a daily newsletter digest; a weekly web radar; a weekly strategy roll-up; and a weekly ideas pipeline. Each runs as a scheduled task — a timed cloud session that briefly connects to the local computer, reads or writes, then disconnects.
Three observations follow:
First, every connection means content from email, calendars and documents — potentially including subject lines and appointments — passes through Anthropic’s processing infrastructure for a time. Even if the output is stored only locally, a cloud model has to see the content sent for processing.
Second, when a routine publishes an output as an Artifact as well as writing it to the wiki, that excerpt becomes a cloud resource. The playbook itself fell into this exact trap: it was created as a cloud document even though it describes a local system.
Third, Claude Memory cuts across both worlds. According to the linked Claude Support documentation, Memory can retain information about previous contexts while chatting. If a project decision is mentioned in a session intended only to maintain a local vault, a derived fact could enter account-linked Memory. That creates another cloud copy — smaller and more abstract than the original, but no longer exclusively local.
What cannot be verified
Intellectual honesty means stating the limits of this analysis as clearly as its findings. Three points remain unresolved:
Admin access on Team and Enterprise plans: whether and how far organisation administrators can access individual users’ content — for example, for audit, compliance or abuse-prevention purposes — depends on contract and configuration. It cannot be checked from an individual user session. Anyone who needs a definitive answer should review their organisation’s contract and admin console.
Usage telemetry beyond content: the public sources reviewed do not explain in detail whether or how metadata such as access frequency, topics or usage patterns can reveal information beyond the content itself. The privacy policy mentions automated processes and internal analysis; that is not a full technical account.
Internal logging practices: content passing through Anthropic’s systems during processing is technically necessary. The public sources reviewed do not describe in detail how long operational logs, error reports or debugging snapshots may exist beyond the stated retention rules.
These are not allegations; they are observations about information gaps. They apply to cloud services generally and are worth addressing explicitly in work involving professional secrecy or regulated data, rather than silently assuming a particular answer.
Practical recommendations
- Choose the storage location deliberately instead of accepting the default. When building a second brain, decide whether knowledge should become a cloud document or a local file.
- Review and deliberately set the model-training preference. Anthropic’s published chat retention period can differ substantially depending on this setting.
- Do not send client secrets or professional-confidential information through routines that publish Artifacts without review. Check whether every automatically published output belongs in a cloud resource.
- Review Memory regularly. If it is enabled, inspect stored facts and remove anything that does not belong there.
- For Team and Enterprise plans, check admin access rights explicitly instead of assuming that “private” means the same thing everywhere.
- Do not accidentally make a local vault’s Git repository public. A private repository is a configuration choice, not an automatic safeguard.
Conclusion
The central point in one sentence: where knowledge ends up depends not only on its substance, but also on the most convenient way to create it — and that path may lead to the cloud by default. A deliberately local second brain stays local as long as outputs are not published as Artifacts and lasting facts are not allowed to flow into account-linked cloud Memory.
This is not a one-time setup task but a recurring decision with every request. For people who work with confidentiality or regulatory duties, it is not a theoretical detail; it is part of the work itself.