How an answer is put together, and what it will not do

One question, step by step

1. The library is searched before anything is asked of a model

   Your question is run against this site's own index — every published essay, and every quotation that has been checked against the archive. The passages it finds are placed in front of the model as its source. This happens server-side and first, so a model that ignored its instructions to look things up would still be answering from retrieved text rather than from memory.

2. The model decides what else it needs, and asks for it

   It can call any of the four tools below — usually two or three times, and often several in one batch. Each call is shown to you as it happens, with the query it wrote, because a reader can judge what was searched in a way they cannot judge a result count.

3. It answers from the passages, and the answer arrives as it is written

   The reasoning models in the chain think before they write, and that thinking is shown separately from the answer — not folded into it, because it is not the answer. The answer streams in under it.

4. Every reference is checked against what was actually read

   A session reference in the answer becomes a link only if a tool really returned it. A reference nothing retrieved stays as plain text, marked as unverified. A fabricated citation therefore shows up as a citation you cannot click — which is a stronger signal than a warning printed beside a working link.

The four tools

Transcribed from the schema the model is given — these are its descriptions, not a summary of them. Three more tools exist in the community's own search CLI and are deliberately absent here: the archive's robots.txt disallows the query parameters they would need, so they are refused by construction rather than merely unused.

search_library

Search this community's own library: every published essay as full text, and every quotation that has been verified against the archive. Returns the paragraph the essay makes its argument in, with the session reference and a citation number. Free and instant - use this first.

required: query

search_material

Search the Law of One sessions at lawofone.info. Use when the library does not cover the question, or when a topic needs the session text itself. Returns passages with session references and a citation number. Keep the query short - one or two words the material itself would use, not a phrase of your own. An empty result means those words are not in the sessions, not that the subject is absent; search a shorter term or the library instead.

required: query

search_channeling

Search the L/L Research channeling library - channelled meditations from 1972 to the present, published at llresearch.org. The Ra sessions end in 1984; these are the same group's continuing work, and they speak most directly to practice: balancing, the disciplines, seeking, and how the circle's own understanding matured. The source varies across the decades - Hatonn, Oxal, Latwii and Laitos in the early years, Q'uo in the later weekly meditations - and each passage carries the source its own transcript names, so attribute a quotation to that source, never by assumption. Returns passages with date references like 2020-09-15.3 and a citation number. Use it when the question is about practice or the present-day work, or when the sessions' wording has not matched.

required: query

read_session

Read a session from the archive in full, as a sequence of questions and answers with their references. Use when a question is about one session specifically.

required: session

list_categories

List the archive's own topic categories, with how many passages each holds. Use to find the material's vocabulary for a subject before searching for it.

list_essays

List the essays this community has written, each with its slug and a one-line summary. Use this to find the right essay before reading it, or to see whether a subject has been written up at all.

get_glossary

Look up the precise meaning of a term of the material - density, catalyst, harvest, the veil, and the rest - as this community defines it, with the session references that ground it. Use when a reader asks what a word means, or before answering with a term the reader has not used. Only a few terms have a written definition; when one does not, this returns the library passages that discuss it and says so, which is still useful.

required: term

read_essay

Read one of this community's essays in full - the careful written treatment of a subject, where a search snippet is not enough. Use when the question needs the whole argument rather than a passage.

required: slug

read_reference

Read one passage by its session reference, such as 17.30 or 74.11, without pulling the whole session. Use when the answer turns on what one specific exchange says, or to check a citation before quoting it.

required: ref

search_memory

Search the notes this reader's own browser keeps about their study: the questions they have asked here, the sessions they have read, the subjects they return to. Use when the question refers to something earlier, or when the notes at the end of your instructions look like a selection rather than the whole picture - which they usually are. Returns matching lines. These are the reader's own records and not passages of the material: quote them only as notes, never with a session reference, and never cite them by number.

required: query

remember

Propose one line for this reader's own notes, so a later sitting can continue where this one left off. Use at most once per turn, and only for something about their study that will still be true in a month: a thread they are following, a subject they found hard, a question they left open. Write it as one short line in plain words, and never write anything personal about them - nothing about their life, health, relationships, or inner state.

required: line

The bounds it works inside

Up to 20 retrieval rounds per question

A ceiling that exists because a model handed fewer tool calls than it asked for does not report an error — it asks again, forever. A malformed-transcript check is the other stop on that failure. Reaching the ceiling is not a failure: the agent still reads its findings and answers from them, and only the fact that it stopped by budget rather than by its own judgement is reported.

100k tokens of working context

Past that, the session folds. Your own questions are kept word for word; the agent's reasoning is kept for the last 5 turns and folded into a record before that — a record of things that happened (queries run, passages read), never a summary written about them, because a summary can invent and a record cannot.

One case is different, and it is worth stating rather than discovering. If your own questions alone fill the whole window, the oldest of them stop being sent as well. Something has to give at that point, and the alternative would be summarising your words into a version you never asked — a paraphrase of a question is a different question, and the agent would answer that one. So the oldest are archived verbatim: still in the record, still in your words, no longer re-sent. The transcript says when it happens, and the record marks which ones.

12 questions every five minutes

A guard on a public endpoint that spends someone's account. It exists because the feature is free to run: with no bill to arrive, nothing else would make sustained abuse visible.

Nothing held between sessions

The transcript lives in this browser tab and nowhere else. There is no account, no history, and no cookie — reloading the page ends the session, and that is a design decision rather than a limitation.