korchasa/save

Memex Save

Save a new source into the memex (long-term knowledge bank for AI agents) and integrate it. One source typically touches 5–15 memex pages.

When to invoke

Invoked when the user wants to capture a source into the memex (e.g., "save this article to my memex", "add this URL to the memex", /save <path-or-url-or-text>). The argument is the source to read.

Memex Resolution

Resolve the active memex root in this order — first match wins:

1. --memex <path> flag in arguments → use that path.
2. Walk up from the current working directory. The first ancestor that contains BOTH AGENTS.md and a pages/ subdirectory is the active memex root.
3. If nothing matches, fall back to ./ in the current working directory and scaffold a memex there (see "Scaffold a new memex" below).

After resolution, read <memex-root>/AGENTS.md for the schema. All conventions (filenames, frontmatter, link format, log format) come from there. Cross-references use SALP only: [REF:mx-<type>:<slug>] (bare) or [REF:mx-<type>:<slug> | <display>] (with display text). [[wikilinks]] are no longer recognised by the audit script or the skills.

The pages directory is pages/ — it holds the cross-linked page graph. The umbrella concept (sources + pages + log + schema) is the memex. SALP REFs ([REF:mx-<type>:<slug>]) are the canonical cross-link notation; do not confuse them with directory naming.

Scaffold a new memex

If no memex was found and the user did not pass --memex, scaffold one in the current working directory:

1. Create ./pages/, ./pages/answers/, ./raw/articles/, ./raw/attachments/.
2. Copy the schema asset to ./AGENTS.md. The asset ships with this skill at assets/AGENTS.md in the framework. If the user already has AGENTS.md in this directory, do NOT overwrite — append a "## Memex Schema" section that links to it instead, then create a separate MEMEX.md with the schema. Ask the user once which they prefer if both options are unclear.
3. Create ./pages/index.md with the empty-catalog template:

   # Memex Index
   
   Last updated: <today>
   
   <!-- Add entries after each save. Format:
   ## <Domain>
   - [REF:mx-concept:page-slug | Page Slug] — one-line description (YYYY-MM-DD)
   -->
   

4. Create ./log.md with the header:

   # Memex Activity Log
   
   <!-- Append-only. Never edit existing entries. Format:
   ## [YYYY-MM-DD] <op> | <title>
   <one-line description>
   -->
   

5. Print: Scaffolded new memex at <memex-root>. Schema is in AGENTS.md.

Continue with the save steps below in the same invocation.

Save Steps

1. Acquire the source

• URL (http://, https://): use WebFetch with the prompt: Extract the complete article content. Return: title, author(s) if listed, date if listed, and full article text preserving all factual claims, data points, code examples. Format as clean markdown.
• File path (existing file): read directly via Read.
• Free text: the user passes the body inline. Derive a title from the first heading or first sentence; if neither works, ask the user for a title.

If a URL is paywalled or returns thin content, report the failure and suggest pasting the text manually. Do NOT fabricate content.

2. Classify and slug

Classify as one of: article | paper | transcript | conversation | note. Auto-detect from the URL or content (arxiv → paper, github → repo-style article, twitter/x → conversation, etc.).

Generate the slug:

• Lowercase the title, replace spaces with hyphens, strip non-[a-z0-9-] characters.
• Truncate to 60 characters.
• Prefix with today's date: YYYY-MM-DD-<slug>.md.
• If the file already exists, append -2, -3, etc.

3. Save to raw/articles/

Write to <memex-root>/raw/articles/YYYY-MM-DD-<slug>.md with frontmatter:

---
date: YYYY-MM-DD
source-url: <URL or original file path or "MANUAL">
source-type: <classification>
title: <extracted or inferred title>
---

Followed by the full source content.

raw/ is immutable — once written, never modify these files in later operations.

4. Extract entities

Read your own raw save. Identify mentioned entities:

• People — named individuals who appear with a role.
• Concepts — ideas, technologies, events, frameworks, theories worth their own page.
• Existing memex pages that this source touches.

For each entity decide: CREATE (no page yet) or UPDATE (page exists in <memex-root>/pages/). Skip minor mentions — they will be referenced inline within other pages, not turned into pages of their own.

5. Write or update entity pages

For each new entity, write <memex-root>/pages/<slug>.md using the matching template from AGENTS.md (concept / person / source-summary). Fill the frontmatter, write a synthesized body (do NOT copy-paste raw content — explain in your own words), and add SALP REFs ([REF:mx-<type>:<slug>]) to related entities.

For each existing entity, read the page first, then integrate new information using Edit:

• Preserve all existing SALP REFs.
• Add new SALP REFs to other entities the new source mentions.
• Add the new source to the page's "Sources" section.
• Update date in frontmatter to today.

6. Write the source-summary page

Always write a source-summary page at <memex-root>/pages/<source-slug>.md covering the source itself:

• One-paragraph summary.
• "Key Points" — 3–7 bullet points.
• "Entities Mentioned" — SALP REFs ([REF:mx-<type>:<slug>]) to every entity page touched in step 5.

This gives every raw source a discoverable memex entry point.

7. Backlink audit (CRITICAL — do not skip)

For each NEWLY CREATED page (from step 5–6), grep all existing memex pages for mentions of its title or key terms:

grep -rln "<new page title>" <memex-root>/pages/

For every match that mentions the title in prose but does NOT contain [REF:mx-<type>:<new-slug>], add a SALP REF at the first natural mention. This is the most commonly skipped step — without it, the memex does not compound.

8. Update pages/index.md

Open <memex-root>/pages/index.md. For each created or updated page, ensure it has a row under the appropriate domain heading:

## <Domain>
- [REF:mx-concept:page-slug | Page Slug] — one-line description (YYYY-MM-DD)

Update "Last updated" to today. Keep entries under 80 characters. Group by domain — derive from tags frontmatter.

9. Append log.md

Append one entry to <memex-root>/log.md:

## [YYYY-MM-DD] save | <source title>
Stored <type> from <source>. Created N pages, updated M. Touched: <slug-1>, <slug-2>, ...

10. Report to user

Print a concise summary:

• Source slug saved to raw/articles/.
• Pages created (with paths) and updated (with paths).
• Index domain headings touched.
• Number of backlinks added during audit.
• Any contradictions flagged with [!WARNING] callouts.

Contradiction Handling

When a new source contradicts existing memex content, do NOT silently overwrite. Add an Obsidian-style callout near the contradicting claim:

> [!WARNING] Contradiction with [REF:mx-concept:other-page]
> Source A claims X, but [REF:mx-concept:other-page] states Y. Needs human resolution.

Touch BOTH pages with the warning, and mention the contradiction in the log entry. The audit skill (audit) surfaces all [!WARNING] markers in its report.

Error Handling

• WebFetch failure (paywall, auth wall, network): retry once. If still failing, report and stop. Do not fabricate the source content.
• Source already saved (matching slug exists in raw/articles/): tell the user and stop. Suggest they pass a different source or remove the existing file first.
• Source too large (>50 KB raw): warn the user that entity extraction may be incomplete; proceed.
• raw/articles/ does not exist in the resolved memex: create it. The memex is malformed but recoverable.
• pages/index.md missing: recreate it from the empty template before step 8.

Constraints

• Never edit files in raw/ after the initial save in step 3.
• Never use plain markdown links for internal references — SALP REFs ([REF:mx-<type>:<slug>]) only.
• Never invent facts. Every claim in a memex page must trace back to a raw/ source via the Sources section.
• Use Read, Write, Edit, Glob, Grep, Bash (for grep only — no git operations unless the user asks). WebFetch only for URL sources.