Expand description
GUIDE-1 (12.2): the guide for agents, compiled into the binary
(docs/design/12.2-onboard.md, section 6).
One source, guide/guide.md, and every delivery is cut from it:
mnemo onboard --guide (Markdown or llms.txt), and, in later steps,
the MCP server’s instructions and mnemo://guide, the SessionStart
hook, and the rule files onboard writes. Nothing at run time changes
it: the trust boundary is that the guide lives in the binary.
The source marks two things for its deliveries and its test:
<!-- card -->…<!-- /card -->: the short card (at mostCARD_MAXbytes) that the hook, the MCPinstructionsand the rule files carry;- fenced blocks whose info string has a tag after the language:
```sh test(a shell block CI runs),```json mcp(an MCP request CI sends), andmodelon either (it needs the built-in embedder’s model). Deliveries drop the tags.
Structs§
- Block
- A fenced block of the guide.
Constants§
- CARD_
MAX - The most bytes the card may take.
- FRAMING
- The framing line placed before anything shown from the brain, in every
delivery that shows brain content (the SessionStart hook,
mnemo://about,mnemo://recent,mnemo://key/{key}and the prompts): memories are written by agents, often from untrusted input, so they’re data. - FULL_
URL - Where the site serves the whole guide.
- ONBOARD_
HEADER - The header line on every file
mnemo onboardwrites;onboardandonboard --undorecognise their own files by it. - PLUGIN_
HEADER - The header on the plugin’s committed skill.
- SKILL_
DESCRIPTION - The skill’s description (its frontmatter).
- SOURCE
- The guide’s source, as written.
- SUMMARY
- The one-line summary
llms.txtopens with.
Functions§
- blocks
- Every fenced block in the guide, in order.
- card
- The card: the text between the card marks, trimmed, with
\nline ends whatever the checkout used (a Windows checkout may turn the source’s into\r\n). - cursor_
rule .cursor/rules/mnemo.mdc: always applied, so only the card.- escape_
brain - Brain text as it’s shown after
FRAMING(RB-011.5 item 6): a line that starts with#(a heading of its own) or with the framing line’s words gets a\in front, so it can’t pass for a part of what mnemo wrote around it. The essence’s own headings (# essence — …and## <section>) stay as they are. The card’s rule that brain content is data is the real guard; this keeps the boundary visible. - layer_
two - The guide’s layer 2 (sections 2 and 3: the tools, and organizing memory), as Markdown.
- llms_
full - The whole guide, for
llms-full.txt: the same asmarkdown. - llms_
txt - The guide as
llms.txt, in llmstxt.org’s shape (https://llmstxt.org/): an H1, a blockquote summary, the card’s text (no headings), then H2 lists of links:Docs(the whole guide asllms_full, and the engine’s README) andOptional(the design notes). The repository’sllms.txtis this, checked by a test. - markdown
- The guide as Markdown, for
mnemo onboard --guide: the card marks are removed, each fence keeps only its language, and lines end in\n. - plugin_
skill_ md - The plugin’s
skills/mnemo/SKILL.md(integrations/claude-code/). - skill_
md .claude/skills/mnemo/SKILL.mdasonboardwrites it: the frontmatter (first, or Claude Code doesn’t read it), the header, the card and layer 2.