Aletheia33. Facts
Facts are the atomic, declarative pieces of memory that sit under summaries, search and the session-start context. They are extracted from captured memory, never from a model's live output.
33.1 Eligible sources
Only two memory kinds are fact sources:
| Kind | Meaning |
|---|---|
summary | An assistant's session summary, captured by a Stop/SessionEnd hook. |
note | A note saved by a person (UI, CLI or MCP). |
Everything else is deliberately not a fact source. A prompt is an instruction or a request, not a fact about the world, and an event is tool output or file contents — it often carries line-number prefixes, code, and other noise that produces junk facts. Raw back-filled transcripts are excluded for the same reason.
33.2 Quality gate
Before a candidate fact is stored it must pass this gate. The first failing check is the reason code.
| Reason | Rule |
|---|---|
unbalanced_quotes | An even number of " and of ' in the whole text. |
unbalanced_brackets | Matching counts for ( and ), and for [ and ]. |
escaped_whitespace | No literal \n, \t or \r. |
code_token | No $, ->, ::, =>, braces, backticks, HTML tags, or uppercase SQL keywords. |
narration | The text is about the project, not the conversation or a moment: no first/second person (I, we, you, your, my), no now, today, yet, currently or right now, no progress/status phrases (is running, are running, is done, are done, gone, in progress, waiting on, will be, has been, was created, not allowed, denied), no this session, the owner, tell me or next session, and no imperative opening (Do, Don't, Run, Open, Merge, Publish, Wait). |
word_count | 3–40 words in the whole text. |
char_count | At most 200 characters in the whole text. |
subject_words | A subject of 1–8 words. |
subject_continuation | The subject does not start with and, but, then, how, to, do or not. |
subject_pronoun | The subject does not start with a bare pronoun/demonstrative such as There, It, This or Everything. |
object_mid_clause | The object does not end on a dangling conjunction, preposition, article, auxiliary or relative word. |
confidence_floor | The stored confidence is at least facts.minConfidence. |
The confidence calculation is unchanged; the floor only blocks storage below the setting.
The rule-based extractor splits a source into sentences before applying this gate. A dot is a sentence boundary only when it is followed by whitespace or the end of the text, so a dot inside a version number (20.19.6), an IP address (127.0.0.1), a file name (config.toml), a domain (example.com) or a URL stays in its sentence instead of fragmenting it.
33.3 Confidence floor
facts.minConfidence is stored in the Hub settings table under ui.settings.facts.minConfidence. The range is 0.3 to 0.9 and the default is 0.4. The rule-based extractor stores facts at 0.45 (0.35 when the object is inferred), so a floor above 0.45 stops it from storing any fact; raise the floor only with a model provider. It can be changed through the settings UI, GET/PUT /api/hub/settings, or directly in the settings table.
33.4 Cleaning existing facts
aletheia facts clean # dry run (default): counts and the first 10 examples per reason
aletheia facts clean --apply # dismiss each failing fact with reason quality:<reason>
aletheia facts clean --ineligible-sources # also reject facts whose source episode kind is not summary or note
aletheia facts clean --apply --ineligible-sources--apply uses the normal dismissal path, so every row is reversible with aletheia restore fact:<id>. Nothing is deleted. --ineligible-sources adds the reason ineligible_source for a fact whose source episode kind is not summary or note; without the flag those facts are left alone.
33.5 Session-start context
The session-start Facts section already skips dismissed facts. Each fact line is capped at 140 characters so a single long fact cannot dominate the context.
The context.facts setting controls which facts the section lists, and is stored under ui.settings.context.facts:
| Value | Behaviour |
|---|---|
auto (default) | List a fact only when it is pinned (its owning observation is pinned), its extractor mode is provider, or its confidence is at least 0.6. The rule extractor's 0.45/0.35 output is left out. |
on | Keep the historical behaviour: list up to 10 active facts. |
off | Hide the Facts section entirely. |
It can be changed through the settings UI, GET/PUT /api/hub/settings, or directly in the settings table. The measured context-economics line counts only what is actually listed.