LLM Wiki
The trustworthy core of your memory: a curated knowledge base of who you are, built from your own notes, with a source on every entry and answers that can only cite what they actually read.
Free-form search over your chats is powerful, but it can surface the wrong snippet or read too much into a passing comment. For the facts about you that must hold up — your profile, your preferences, the people and projects you care about — the Personal Brain keeps a separate, curated layer: the LLM Wiki. It is the part of your memory built to be reliable, not just retrievable.
The wiki is built directly from your own curated notes — there is no AI "fact extraction" step that could invent or distort what you meant. Every entry records where it came from, and when an agent answers, it can only cite entries it actually opened. It cannot make one up.
A knowledge base of you#
The wiki is a hierarchy of knowledge entries, organized like a set of folders. Each entry has an address, so it can be browsed and linked, for example wiki://you/preferences/database. Entries are grouped into six categories:
Who you are: role, background, and the basics that rarely change.
How you like things done: tools, style, defaults, and standing choices.
The people, teams, and projects you work with.
Notable moments in your history, kept as durable knowledge.
Worked examples and decisions you can reuse later.
Habits and rules of thumb the brain has learned about you.
Three levels of detail#
Every entry is written at three levels of detail, from a one-line abstract to the full text. An agent reads the cheap summary first and only opens the deeper level when the entry turns out to be relevant. This keeps answers fast, focused, and easy to audit.
Why you can trust it#
The wiki is more trustworthy than free-form search because credibility is built into how it is made and how it answers:
- Built from your own notes, not guessed. Entries come from your curated memory by a deterministic process. There is no model step that infers facts about you, so the wiki cannot invent or drift from what you actually wrote.
- A source on every entry. Each entry records where it came from. Re-run the build and entries are replaced from their source, so nothing stale lingers.
- Answers are citation-locked. An agent may answer only from entries it has read this turn, and the citations it returns are filtered to exactly those entries. It cannot cite — or fabricate — something it never opened.
- Curated summaries, not random chunks. Search returns vetted entry abstracts, so relevance is judged over clean summaries instead of arbitrary text fragments.
- Auditable retrieval. Every step — what was searched, read, and cited — is recorded, so any answer can be replayed and checked.
- Read-only recall, isolated writes. Answering never changes your wiki. Updates happen only through the build step, all-or-nothing, and always scoped to you.
If your wiki does not contain the answer, the agent says so with low confidence rather than guessing from general knowledge. A clear "I don't have that" is more useful than a confident wrong answer.
How it's built#
You build your wiki from memory you already keep — your own profile and notes about how you work. The build is deterministic: your notes are parsed and mapped into entries, preserving their structure and category. Because there is no inference step, the same notes always produce the same wiki, and no fact appears that you did not write.
Point the build at your notes: your profile and the running memory your agents keep about you.
Each note becomes an entry at its category and level of detail. The mapping is structural, so what goes in is what you wrote.
Notes edited on more than one device are merged line by line, honoring edits and deletions, so a rebuild never silently loses anything.
How it's read#
When an agent asks something the wiki can answer, it works through the entries like a person browsing a well-organized notebook:
| Step | What happens |
|---|---|
| Browse | Navigate the hierarchy of entries like folders to get oriented. |
| Search | Find the entries most relevant to the question and read their abstracts. |
| Read | Open the overview, then the full text, only for entries that look relevant — up to a strict limit. |
| Answer | Return a grounded answer with a confidence score and citations to the exact entries it read. |
Wiki, memory, and graph#
The three concepts fit together, each doing what it is best at:
Your timeline: everything you did and said, kept in order.
The links: how your data connects to the concepts you use.
The trusted core: curated, sourced facts about who you are.
When you ask a question, the brain draws on your memory and your wiki at once and merges the result, so you get both the timeline of what happened and the trusted facts about you — in one grounded, cited answer.
How every agent and device you use shares this one memory.