Your AI Second Brain Should Remember When You Changed Your Mind

This week, the concept of "second-brain lists" resurfaced on X, with notable examples including Karpathy's LLM Wiki, Garry Tan's GBrain, and various Claude-plus-Obsidian integrations. The key takeaway is the significance of persistence in AI second brains - not only in retaining facts but also in remembering when perspectives change.

An article highlighted in the feed emphasized that an AI second brain should not only remember preferences but also be capable of recognizing when to retire them, as personalization becomes increasingly based on an outdated portrayal of the user. Failure to address this issue is evident in team wikis, where decisions made in previous quarters are still cited, even when those decisions were later reversed.

The crux of the problem lies not in amnesia, but in the inability to retire outdated truths. A larger context window does not solve this issue; it merely stores more information without expiring any. Similarly, advanced Retrieval-Augmented Generation (RAG) techniques also fail to address the problem, as they tend to surface the most recent decisions rather than the most recently reversed ones.

The solution lies in effective context engineering, which involves distinguishing between current, superseded, and retired information. This distinction is crucial in maintaining a coherent knowledge base. For instance, consider a stack decision where an RFC from June states the use of an 8B model in production, accompanied by an evaluation and cost cap.

The LLM wiki reflects this decision with a concise page indicating the production model as 8B, citing RFC-142. Later, in October, the team switches to a larger model due to a latency incident, leading to an updated RFC or a new one being filed. However, the wiki page remains unchanged, reflecting the earlier decision. Consequently, when an agent provides an answer based on the June page, it inadvertently reinforces the outdated decision.

To resolve this, important pages should be treated more like interfaces rather than mere souvenirs. Each page should have a status (current, superseded, or retired), an "as of" date, a review date, and a pointer to the document that replaced it, rather than a silent rewrite. The owner of the page should also be identifiable to confirm its continued relevance.

Raw sources should remain immutable, while synthesized wiki pages are allowed to change, but any change of mind should leave a clear trail. The old page should be marked as superseded, and the new page should be what agents retrieve first. In the case of schema files, such as AGENTS.md or CLAUDE.md, it is essential to explicitly state that preference should be given to pages with status: current.

Conflicting pages should be addressed by determining which one is currently in force, rather than attempting to average them. When a human updates a decision, the corresponding wiki summary must also be updated in the same change, ensuring that agents retrieve the most current information. While a weekend folder of agent-written notes might seem sufficient for individual use, it falls short of providing a comprehensive solution for team knowledge management.

Knowledge base software, which already exists and is widely used, is required to effectively implement this system. Such software should include permissions, page history, and a clear status indicator within the open environment. By integrating AI knowledge base software, teams can ensure that agents only retrieve relevant, up-to-date information, eliminating the issue of stale knowledge.