Your AI Second Brain Should Remember When You Changed Your Mind

This week on X, the second-brain lists are back again: Karpathy's LLM Wiki, Garry Tan's GBrain, a pile of Claude-plus-Obsidian hooks. The pitch is persistence. Don't reload context every chat. Keep the facts.

That is necessary. It is not sufficient.

A sharper point showed up in the same feed: an AI second brain should remember when you changed your mind. Otherwise "personalisation" becomes an increasingly confident portrait of someone you used to be. Remembering a preference is easy to demonstrate. Knowing when to retire it is the actual product problem.

The same failure shows up in team wikis. Last quarter you chose a vendor, pinned a model, and banned a library. The decision was real. Then you reversed it. The wiki still cites the old page, because the old page is well-linked, well-written, and still technically true — as of September.

The bug is not amnesia. It is unretired truth.

A bigger context window does not fix this. AI can now hold millions of tokens and still fail to connect which paragraph is still in force. A bigger window stores more. It does not expire anything.

Over-engineered RAG does not fix it either. Retrieval will happily surface the most similar, most linked, most frequently cited decision. That is often the first decision, not the latest one. The useful work is context engineering: what is current, what is superseded, and what an agent is allowed to treat as canonical.

This is a different failure from meaning drift. Drift invents a more general claim than the source supported. Stale knowledge repeats a claim that was correct, cited, and later reversed — and the reversal never made it into the pages the agent reads.

What "changed our mind" looks like in a wiki

Take a stack decision.

1. In June, an RFC says: "We will serve the 8B model in production. Record the eval and the cost cap."
2. The LLM wiki writes a crisp page: "Production model: 8B. See RFC-142."
3. In October, you switch to a larger model after a latency incident. The RFC is updated, or a new one is filed. The wiki page is not.
4. In November, an agent answers an on-call question from the June page, with a tidy citation.

Every fact kept its source. The source was just no longer the ruling document.

Personal vaults hide this because you remember the reversal. Teams do not. The person who changed the decision is on leave. The agent only sees the pages.

Knowledge that can expire

Treat important pages like interfaces, not souvenirs.

Each canonical page needs:

• A status: current, superseded, or retired
• An "as of" date, and a review date
• A pointer to the document that replaced it, not a silent rewrite
• An owner who can say the page is still in force

Raw sources stay immutable. Synthesized wiki pages are allowed to change, but a change of mind should leave a trail: the old page stays, marked superseded, and the new page is what agents retrieve first.

If you are writing a schema file (AGENTS.md, CLAUDE.md, or the equivalent), make this explicit:

• Prefer pages with status: current
• Never cite a superseded page as the answer; cite it only as history
• If two pages conflict, ask which one is in force, then stop — do not average them
• When a human updates a decision, the wiki summary must be updated in the same change, or the change is incomplete

Clean markdown with those fields beats a more elaborate pipeline that cannot tell June from October.

Where this belongs

A weekend folder of agent-written notes will not hold this for a team. You need Knowledge base Software people already edit, with permissions, page history, and a place for status to live in the open.

If agents can search and cite those pages, you have AI knowledge base Software. If you can run the workspace yourself — the thing people usually want from an Open source Notion alternative — you also get to decide whether inference leaves the building.

AFFiNE's write-up on an LLM wiki for teams keeps raw sources read-only and treats the wiki as a maintained layer on top, with agents starting read-only over MCP. The companion guide to self-hosted AI for teams is useful here for a different reason: the knowledge workspace is its own layer. You can expire a decision in the workspace without retraining a model, and you can swap the model without losing the changelog.

AFFiNE, which I work on, is one option for that workspace layer: open-source, local-first docs, whiteboards, and databases, with optional BYOK AI and a workspace MCP server that is read-only by default. It is a weak fit if you only need a static handbook, or if you need agents to write into the workspace on today's stable release.

A one-hour stale-knowledge check

Pick a topic your agents already answer: default model, pricing, auth, or "what we use for X."

1. Ask the same question twice: "What is the current decision?" and "When did we last change our mind?"
2. List every page or chunk the answer cites. Mark each one current, superseded, or unknown.
3. Find the human document that actually governs the decision (RFC, ticket, changelog). Check the dates.
4. If the wiki page is older than the governing document, mark it superseded and write a current page that links to both.
5. Add status and review dates to the schema your agent already follows.
6. Ask the two questions again. The first answer should cite the current page. The second should cite the superseded page as history, not as policy.

If the agent cannot answer the second question at all, you do not have a memory system. You have a pile of still-sounding facts.

Bottom line

Persistence without expiry is how a second brain becomes a museum.

Let the LLM wiki keep the history. Just do not let last quarter's decision keep answering this week's questions. A useful second brain is not the one that remembers everything. It is the one that knows what you no longer believe.