Using Fractaloom

Fractaloom is your organization's memory: it reads the systems your administrator connected (GitHub, Slack, Google Drive, Notion, Jira) and answers questions with evidence. It only ever shows you what you could already open in those systems. This guide is for everyone who asks it questions; the pages it describes are the entries in the sidebar, grouped by what you do there: Home and Ask at the top, then Explore (Search, Browse, Entities, Decisions, Timeline, Brain), Work (Research, Spaces, Packets, Standing questions, Agents), Add knowledge (Capture, Memory, Sources) and Keep it right (Knowledge Health), with Help below.

Ask a question

Open Ask, type the question the way you would ask a colleague ("Why did we choose PostgreSQL?", "Who owns the payments service?", "What is our current Enterprise price?") and press Enter. The answer is written while you watch, with numbered citations such as [f1] that link to the evidence below it: the documents, messages and tickets the answer rests on, each with a link back to where it lives.

Under the answer you find:

Current state: the facts the organization holds about the subject right now, each with its evidence.
Historical context: what was true before and when it changed.

You can also ask how things get done: "How do incidents get resolved here?" or "How do we handle renewals?" answers with the steps the organization actually follows, who usually acts at each and how long each takes, learned from the incidents or renewals you can see and cited to them. It needs at least three of them with their events; with fewer the answer says so.

Decisions: what was decided about the subject, with the rationale and the alternatives that were rejected.
Related entities and experts: who and what is connected, and who to ask.
Caveats: anything you should know before relying on the answer (an outdated source, a conflict, a date cutoff).

A citation from an earlier version. When a document has changed since an answer's source was read, the evidence says "Quoted from the version of 3 September; the document has changed since" and Read that version opens the text it quoted, above the current one. What a document no longer says is history from the day it changed, and still the answer when you ask about an earlier date.

Ask about a date. The date box next to the question answers as of that day: "What was the Enterprise price on 1 March?" gives what was true then, not now. Leave it empty for today.

Ask inside a space. Pick a Context Space in the dropdown to bound the answer to one project, team or customer (see Spaces below).

Starter questions. Home suggests questions the organization can answer well right now; click one to ask it.

File as note. A good answer should not vanish in a chat. Under the answer, File as note keeps it in your own notes or in a team's memory, with the sources it showed numbered and each citation in the text turned into the matching number, and the note is read like any other document: it turns up in search, in later answers and in your coding agent's recall. The people in a team see only what you file to that team.

Ask about a group. "Which renewals got a discount this year?", "What keeps causing our incidents?", "What do our churned customers have in common?": the answer says how many there are that you may see, what recurs across them with how many share it ("a discount off list price: 2 of 3"), and then each of them with its sources, folded under Themes on the answer. When nothing recurs, it says so. You only see the members, and the lines about them, that you may read.

Read the badges: knowledge state and confidence

Every answer carries two badges. They are the product being honest about what it knows.

Observed: stated directly in a source.
Verified: confirmed by an authoritative source or by a person who may verify it.
Inferred: derived from several signals rather than stated outright.
Historical: was true, is no longer current.
Conflicted: reliable sources disagree; the answer shows both sides.
Unknown: not enough evidence. The answer says so instead of guessing.

Confidence is low, medium or high. Hover or click the badge for the four factors behind it: how much of the question the evidence covers, how settled the facts are, how many independent sources agree, and how recent the evidence is. A low-confidence answer is not wrong; it is thinly supported, and the evidence list shows why.

Every badge explains itself where it is: hover a knowledge-state badge for one sentence on what the state means, click the confidence badge for the factors. The rest of the page speaks in words too: a fact is "well supported" or "thinly supported" and "observed 3 weeks ago, still fresh", a question was "asked as a current fact", a routine runs "weekly". The numbers behind the words are in the tooltip of each phrase and under Trace; nothing is hidden, it is just not the first thing you read.

An unknown answer is a feature: nobody wrote the answer down where Fractaloom can see it, or you may not see the document that holds it. Both look the same to you by design.

An answer written from a passage rather than an extracted fact (a note, an incident review, a runbook, a pull request) says so in its caveats: "Answered from passages". It is still cited and still visible to you only if the document is; its confidence is lower because no fact was extracted and checked. When you ask about something nobody recorded for a subject (an NPS score, an owner that was never named) the answer says exactly that instead of listing the subject's other facts. Inside a Space, a question the Space does not cover answers "Not in this space" with what the Space holds; ask again without the Space to search everything you may see. A proposal that was never decided ("Introduce a Team tier at $49") is said as one when you ask about it: "Proposed, not decided", with who proposed it, and only if you may read where it was proposed.

When an answer is wrong

Under every answer: Was this right? Click No, it was wrong, say what was wrong and what is right, and send. The note goes to the people running the beta, and the answer's sources are marked; a source that keeps misleading answers is flagged for the administrators under Knowledge Health. Yes and opening a source from an answer teach the ranking too: sources behind right answers and the ones people open rise for your organization, sources behind wrong answers fall.

If a fact itself is wrong, fix it where you read it: every fact in the answer's Current state and on its entity page carries correct and retract for members (see Entities below). If something else went wrong, Report a problem in the sidebar foot opens the support address with the id of your last answer filled in, so the team can find exactly what happened. The same id is shown under the answer as request.

Your inbox

Home shows an inbox when something waits for you: a fact you own that is due for verification (still true or no longer true settle it), your weekly note (the counts of Your week), and, for administrators, the weekly lint note: what the memory needs next, with a link to Knowledge Health → Proposals where each line is one click from being settled.

Your week

Your week on Home lists what the brain learned about your work in the last seven days (your projects, customers, systems and teams, what they are linked to and the Spaces you own), each line with its source; what waits for you (a decision to confirm, a fact to verify, a decision to review, an approval, sources that disagree); and, when you connected a source or administer them, the sources it could not read. You only ever see what you may read, and it is put together when you open Home. Once a week your inbox gets a short note with the counts; turn it off under Send me a weekly note on your Account page. Your coding agent can ask for the same with the my_week tool.

Entities: the organization's objects

Entities lists the people, projects, systems, customers, products and policies Fractaloom knows. Open one for its page: current facts with evidence, history, decisions, relationships, the people who know it, and a timeline. In brief at the top is the page as prose, written weekly for the busiest projects, customers, products, services and teams and refreshable by any member: what changed lately, the open issues, the decisions that shape it and the facts to know, each line linked to its source; as with a space's digest, you see the lines whose sources you may read.

On a fact a member has four controls. verify: you confirm the fact is right with you as the evidence, and it stays verified until a source contradicts it or its owner is asked to check it again (owners get a reminder when a fact is due for verification). correct: type the right value, say what happened and, if you like, why; a new verified fact with you as its evidence replaces the old one. It was wrong (the default) means the old value never held: it leaves the history, and the source that said it counts as misleading. It changed on a date means the old value was right until then: it stays in the history until that day, the new value holds from it, and no source is blamed. Knowledge Health → Conflicts asks the same when you mark one value correct: the others were wrong, or held until it changed on a date. retract: the fact leaves the truth (no answer will use it) after you confirm; it stays listed under Retracted facts with undo retraction, which puts it back exactly as it was. Every fact links to its history: what it replaced, what replaced it, and every correction with who made it, when, why and the documents it overrode. Your coding agent's report_correction tool does the same from the terminal. set a review: pick who checks the fact and how often (every 90 days, say); they get a reminder when it is due, and a page whose owner or status fact has a review date that has not passed reads Canonical. Tick test on a correction you make only to try the workflow: it takes effect like any other and is marked as a test in the history, but it never counts against the source or the extractor behind the fact.

add another name under an entity's name records what people call it ("Phoenix", "the ledger rewrite"), so questions that use that name find the entity; anyone who may see the entity sees the name. Corrections are visible to everyone who can see the fact, and real ones count against the extraction that produced the wrong fact, so the same mistake is less likely next time. The same goes for No, it was wrong under an answer: the sources the answer cited, and only those, are marked.

Your department first. Home shows your department's latest decisions and the ones you own or approve, the proposals waiting for you to confirm, and beside each the closest thing done before. The Decisions page lists your department's decisions first (pick "My department only" to see nothing else), and on Ask you can name a department: "What has sales decided recently?" or "Which decisions did engineering make about Phoenix?" answers from that department's decisions only and says so on the answer. Your department comes from your groups; an administrator sets which department each group stands for on Admin → Groups.

An entity's page opens with what it is in one line, the essence (who owns it, its status, the decision that shapes it, what changed last, what is still open, each with its sources), and a badge that says how far to trust it: Draft when part of it is inferred, unconfirmed or contradicted, Validated when its key facts are stated by sources, Canonical when it also has an owner and a review date that has not passed. A line marked [uncertain: …] rests on something no source stated outright; [contradiction: X says …, Y says …] means two sources disagree. Why it is this way lists the decisions behind it with the options they were chosen over, and says which decision a change would reopen and whom to talk to. Related says why each link is there. Compare with lines it up against another of its kind (two plans, two customers) with the decisions behind each difference. A Space's page opens the same way with its centre and its decisions. You might also need, beside an entity's or a decision's page, lists the decisions about it and the earlier ones like it that you may see, each with why it is there and how it turned out ("A decision like this one was taken before"); Home suggests the earlier decision like one waiting for you. Browse in the sidebar walks everything you may see as Projects (things that end), Areas (things kept up), Resources (reference) and Archives (what ended: a shipped project, a signed deal, a resolved incident); something that ends moves to Archives on its own and leaves Home.

Decisions lists what was decided, with rationale, participants and the alternatives that lost; filter it by department or by kind of decision (a pricing exception, a roadmap call, a technology choice…). Have we done this before? at the top of the page takes the decision you face in a sentence ("Acme wants 15% off for a three-year renewal") and returns the earlier decisions most like it that you may see: why each matches, who approved it, whether it still stands and what happened afterwards. Asking "what did we do last time …" or "have we ever …" on Ask does the same. A decision's page says what led to it, in the words of the document (the request, the numbers, the constraint), who approved it, when it is to be looked at again and the earlier decisions it follows. Record a decision on the same page writes one down where it is taken: the decision, why, what led to it, the alternatives and why they lost, its owner and approver, what it affects and when to look at it again. It is filed in one of your teams' memory, so only that team may read it, and it waits as a candidate until its owner or approver presses It was decided on its page (they get a notification when they may read it). A ## Decision section in a pull request's description and "we decided …" said in a meeting transcript become candidates the same way, and a coding agent records one with record_decision. Did it work? on a decision's page shows what happened since it was taken to the things it affected (events, changed facts, issues still open, later decisions), as far as you may see them. When a decision has a revisit date, its owner and approver get a notification on that date; they (or an administrator) press It held, It was revised or It was reversed, with a sentence on what happened, and that verdict appears in Have we done this before? and in answers from then on. The Decisions due for review routine sends you a daily list of the decisions you own or approve that are coming due. What argues against it sits on every decision's page, under each result of Have we done this before?, in Record a decision before you save, and on answers to questions like "What argues against giving Acme 15% off?" or "Why might pausing Nova be a mistake?": an earlier decision like it that was reversed, an open issue or a disagreement between sources on what it touches, a rule or limit that applies, and the options it was chosen over with why they lost. Each point links to its source, and you only see points drawn from what you may read. When nothing recorded argues against a decision, it says so; it never invents an objection. Timeline shows what happened to an entity over time. Search finds documents by words when you want the source rather than an answer. Brain has two tabs. Map shows which regions of the organization's memory are well covered and which are thin. Graph draws the entities you may see and the relationships between them: pick a root with the search box (or arrive from "Show in graph" on an entity page, a Space or an answer), widen the rings, drag the canvas to pan and scroll or use + and − to zoom, hover a node to light what it is connected to, click a node to read about it or an edge to see the sentence that evidenced it. On a phone the canvas is taller and shows part of the picture at a readable size: drag to see the rest, pinch to zoom, and tap a node or an edge to read about it below. Drag a node to move it: the picture settles around it and the node stays where you drop it (Re-layout lets go of every node you moved). The type chips filter the picture and stay visible so you can always clear them, and Back returns to the previous root. Diamonds are decisions, an amber ring is an open issue, a red ring an open conflict. The slider replays the graph as of any month, and the health lens flags what needs attention. You only ever see the part of the graph your permissions allow.

Spaces: a bounded brain

A Context Space is a slice of the organization's memory around a few root entities: a project, a team, a customer. Create one on Spaces with a name, a kind and its root entities; everything within reach of those roots is in the space. Then:

Ask in this space answers only from that slice, which makes answers about a busy project sharper.
Export as skill turns the space into a file a coding agent can carry: the current facts, the decisions with their rationale, the owners. Your administrator's guide for agents explains where it goes.
Digest is the space's standing summary: what changed in the last month, the open issues, the decisions that shape it and the facts a newcomer must know, each line linked to its source. It is rewritten every week, and the space's owner can refresh it any time. You see the lines whose sources you may read; when that is not all of them, the headline says how many you see.

Spaces respect permissions like everything else: a space never shows you more than you could see without it.

Before a meeting

Connect your Google Calendar on Sources (it reads your own calendar only; every meeting becomes a note that you alone can see). Home then shows your next meetings the organization knows something about ("In 2 hours with Acme Corp: Acme Corp renewal sync"); click one and Ask prepares you: what we know about them, open issues, recent decisions and changes, who is coming. The Meeting preparation routine does the same an hour before each meeting and files the preparation in your personal memory, so it is there when you open the call.

Memory: notes that become knowledge

Memory is where you write things down that no system holds: a preference, a decision from a hallway conversation, a convention your team follows. Pick the layer:

Personal: only you see it, and it is yours to delete.
Team: everyone in a group you belong to.
Organizational: everyone in the organization (administrators and members).

The Memory page starts with the note form, then the notes you can see (yours and your teams'), and keeps the layers and their rules (who can read, how long it is kept, who may delete) behind Memory layers and their rules at the bottom; what agents wrote for themselves is listed there under Agent memory and never shows up in your searches or answers unless you ask for the agent layer. When you file an answer as a note, the note's title is offered from the answer's first line and you can change it before filing. A note in a team's memory is deleted by the person who wrote it (or an administrator), never by a teammate, and Delete asks you to confirm.

Save to brain in one click. Capture in the sidebar opens the capture page: an address, a title, the passage or thought to keep and why it matters; it becomes a note here. On your Account page, copy the bookmarklet into a new bookmark; on any page, select a passage (or nothing, to keep the page) and click it: the capture page opens with the address, the title and the selection, you say why it matters and save. In Slack, the message menu's Save to brain does the same for a message; on an answer, File as note takes a why too; a coding agent uses capture. Saying why it matters makes it count for more and is shown with the note. The capture page says when the note is saved, with a link to it, or why it was not, with Try again.

A note is read like any document: it is extracted into facts, cited in answers and shown with you as its source. Each layer has its own retention; a personal note is deleted with you if you are ever forgotten.

Standing questions: one answer the organization keeps

Some questions come up again and again ("What is our current Enterprise pricing policy?", "Who owns each payments component?"). An administrator can press Keep this answer current under an answer, or add the question on Standing questions in the sidebar; a Space's owner can keep one inside their Space. Fractaloom checks the evidence every hour and writes the answer again only when the evidence changed, changing as little as it can, and keeps the history.

When you ask such a question, the answer says Standing answer · version N if the evidence you can see is exactly what it was written from: everyone who sees the same sources gets the same answer, and it comes back faster. If you see different sources, you get a fresh answer from what you may read, never someone else's.

Directives: rules every answer follows

An administrator can set hard rules for the whole organization on Admin → Settings → Directives, and a Space's owner for their Space on its page. Never state keeps what it names (an attribute such as contract.annual_value, or words) out of every answer, its evidence, the context agents read and change reviews; the answer says Withheld by a directive with the rule's label. Always say adds its sentence whenever an answer touches what it names ("Enterprise pricing for Acme is under negotiation"). Everyone sees which rules are in force; only those who may edit a rule see what it names. A directive changes what an answer says, never who may read a source.

What keeps happening

An entity's page shows What keeps happening: a customer that asks for a discount at every renewal, a price that moved three times, incidents that keep coming from the same cause. Each pattern says how many times, since when, whether it is strengthening, steady or weakening (the latest record goes against it, such as a renewal with no discount asked), the quotes it rests on, and its history. It is counted only from the sources you may open, and shown only when you can see it twice in two sources, so two colleagues can see different counts. Ask "Does Bayline keep asking for discounts?" and the answer starts with the pattern.

Packets: pin what you will reuse

A packet is a piece of knowledge you keep with its sources, to use again. Pin it where you read it: a section of an answer (Current state, Decisions, What argues against it), a decision (Pin summary on its page) or a page's essence (Pin essence on an entity's page). Fractaloom writes the packet itself from what you can see, with the source of every line; you never retype it. Packets in the sidebar lists yours and your teams', with how many drafts have used each. A line resting on a source someone may not open is not shown to them, and they are told how many lines they cannot see.

Choose a few packets and Draft a decision memo, an account brief or a proposal section. The draft uses only the chosen packets, and every sentence cites one: each citation links to its source, and a sentence that could not be tied to a packet is removed and counted. File as note keeps the draft in your memory or a team's, with its sources listed.

Copy answer under an answer puts the text on your clipboard with its sources numbered and listed at the end, ready to paste anywhere; Copy on a page's sections does the same for a section. Copies, pins, drafts, opened sources and the decisions a review cites are how Fractaloom knows its answers are used: administrators see the reuse rate and the other measures on Admin → Analytics. Only what was reused is counted, never your text.

Research: a longer question

Deep research is for questions a single answer cannot carry: "What is going on with Project Phoenix?", "How did our pricing evolve and why?". It plans sub-questions, answers each with evidence, groups the findings, points out patterns and conflicts, and writes a cited report you can keep. Bound it to a space when the question is about one project.

Agents, approvals and routines

Agents are assistants with their own name, role and permissions, run by members. Give one a task ("Summarise what changed about the Acme renewal this week and tell #sales"); it answers from the same memory you see and acts only within its execution policy. An action that needs a human (sending a message, writing a note) waits under Approvals until someone approves it; the card says in plain words what would be sent, to whom, and the message itself. A message out is approved by an administrator other than the person who asked (in an organization with one administrator, that administrator may approve their own); you can always reject your own. An agent is kept on task before anyone is asked: a message that says nothing about what you asked is not sent (the agent answers you instead and says why), a hand-off to an agent that does not exist fails at once, and "Remember that …" stores exactly the sentence you gave. Agents are created by administrators; when an agent proposes to turn a repeated task into a routine, it asks which project or customer the routine is about. Routines run an agent on a schedule or when something changes, for example a weekly digest for a project.

Everything an agent did is recorded on its run, with the evidence it used.

An agent also learns from repetition. Ask it the same kind of thing three times ("Summarize the status of Acme", then Globex, then Initech) and it keeps the procedure as a workflow: the Agents page shows what it has learned to do, the agent follows it next time, and your inbox asks whether to make it a routine. Make it a routine creates one you own on the cadence you have been asking at, with the agent's action still waiting for approval under its policy; Dismiss retires it.

Your coding agent

Claude Code, Codex, Cursor and similar agents can ask the same memory from inside the editor over MCP (Model Context Protocol), so "why is this module like this?" gets the decision behind the code. Open Account (your name at the foot of the sidebar), create an API token with a name, an expiry and what it may do (ask, read context, search, administer), and copy it once together with the one-line setup shown under it; the token acts as you, with your permissions and never beyond its scopes, and the same page revokes it. The guide for agents has the details. A pull request can be reviewed against the memory too: the review says when a change contradicts a decision or a current fact. The decisions your agent takes can come back as decisions waiting for you to confirm: ask it to record them, or install the session hook from the guide for agents; only the decisions are sent, never the session.

Files instead of MCP. If your administrator allows it, choose Local mirror or Both under How your agents get context on your Account page: fractaloom mirror pull then keeps a folder of Markdown pages with what you may see (entities, decisions, spaces, and the documents' text if allowed) current on your machine, so your agent reads and greps files; with Both it still writes back over MCP. The page gives you the command and the lines for your CLAUDE.md.

For administrators: fixing the extractor. When people keep correcting one kind of fact (contract terms, owners, prices), Knowledge Health → Misleading extractors proposes a one-line rule for the extraction prompt, at most two a week (or press Propose now). Test reads the corrected documents again with your model without saving anything; Approve is possible only once the test passed, and it reads those documents again with the new version. Admin → Settings shows the version in use and Undo brings the previous one back.

Sources: what a sync produced

Sources are connected by administrators, but any member can open Review what it produced on a source's card on the Sources page. It shows what the last sync read (how many of the documents you may open), what validation dropped or flagged and why, how much was extracted, the facts the extractor was unsure about (a low confidence, or a quote it could not find in the document), the decisions waiting for a person and the documents that contributed most. Confirm or Wrong on a fact and It was decided or Not a decision on a decision are ordinary corrections, recorded with you as the evidence; what you cannot see is not in your review. A coding agent reads the same with source_review. The review also says what was read and not kept, and why (a document with nothing worth keeping, items below the organization's signal bar, items validation dropped, messages kept out by a person's choice), so a quiet source is told apart from a broken one. Not relevant on a fact says it should never have become knowledge: it leaves the answers, and the organization's bar learns from it.

Is the connection working? Each source's card says Connected and tested (a test or the last sync read a document), Connected, nothing to read yet (the connection works and there is nothing to read), or Can't read: reconnect with the reason (for example "api.github.com answered 401: the credential was refused"). Administrators press Test connection to check again; it reads one document and stores nothing. When a sync could not read some documents, the card says how many are waiting to be read again: every sync tries them again first, and administrators can open the list with each error.

Your first question. When the first source has been read, Setup asks the administrator for a question they would normally search for or ask a colleague, answers it and asks Is that right? Say yes, or no with the right answer; the question and your verdict start the organization's benchmark, and Setup keeps showing the question until you have said.

Connect your own mail, calendar and Slack

On the Sources page you can connect your own Google Calendar, Gmail (pick the labels to read; nothing else is) and My Slack activity (what you sent, where you were mentioned, your direct messages). You sign in with your own account; what they bring is yours alone, in your personal memory, and helps the answers you get. Disconnect them from the same card.

What the brain reads

Account → What the brain reads lists every source your organization connected, what it reads (the channels, folders, labels and repositories picked) and who connected it; someone else's own mail, calendar or Slack connection is only counted, never named. There you can keep your own messages out of extraction: they stay searchable to whoever may read them, but no facts, decisions or events come from them, from the next sync. Your personal notes and memory are visible to you only unless you file them to a team.

What you can and cannot do: roles

Viewer: read and ask. The Sources page shows what is connected without the controls.
Member and viewer alike see the knowledge, not the machinery: a source's sync history, the connectors' errors, the internal stores Fractaloom keeps for notes and agents, how an app is registered, and Knowledge Health's buckets about sources, connectors, extractors and permissions are for administrators.
Member: also use Knowledge Health (what conflicts, what is going stale, what nobody owns, with the buttons that settle it), verify and correct facts, write notes, create spaces, run research, agents and routines.
Admin: also connect sources, create agents, manage people, groups, permissions and authority, see analytics and the audit log, and settle identities: when a Slack or GitHub id in the documents is not in the directory, the admin console's Identity tab says who it is, and from then on their mentions, permissions and expertise count for that person.

The full list of what each role can do is at the foot of this Help page and under Add a person in the admin console; both read the same table the API serves, so they never disagree. If a button is missing for you, that is your role; ask an administrator.

What Fractaloom never does

It never shows you a document, fact or decision you could not open in the source system. Permissions are checked on every query, not after the fact.
It never invents an answer: without evidence it says unknown.
It never changes a source system on its own: corrections live in Fractaloom, agent actions wait for approval where the policy says so.
It never keeps a token, key or password it finds in what it reads: it is replaced with a marker such as [REDACTED:github_token] before anything is stored or sent to the model, and card numbers, social security numbers and IBANs too unless your administrator turned that off. If your organization keeps such text out, a note holding one is refused and the message names what it found.

Keyboard and small things

On Ask, press Enter to send; the answer streams in. Reload the page and the question and date stay in the address bar, so a link to an answer can be shared with a colleague who may see the same evidence.
Every list has a filter box; type to narrow it.
Dates are shown as the day; hover for the full timestamp where it matters.
The organization switcher in the sidebar foot appears only when your account belongs to more than one organization (the demo beside yours, say).