| name | knowledge |
| description | Use when saving insights from the current conversation to persistent storage, or when retrieving past insights before starting research or writing tasks. |
Knowledge — Persistent Insight Store
Manages the .agent/knowledge/ folder as a persistent, searchable store of insights gathered across conversations. Two modes: save (write insights) and retrieve (find past insights).
Knowledge Folder
- Path:
.agent/knowledge/
- Format: Standalone
.md files with YAML frontmatter
File Format
Every knowledge file follows this structure:
---
title: "NFW Profile Definition"
date: 2026-03-04
source_skill: literature_research
chapter: chapter_01
tags: [dark-matter, density-profile, nfw, halo]
---
## Summary
[Concise description of the insight]
## Source References
- [Author (Year)] — arXiv: [number], bib key: `[key]`
## Context
[Which chapter/section this was gathered for, and why it matters]
Frontmatter fields:
| Field | Required | Description |
|---|
title | Yes | Descriptive title of the insight |
date | Yes | Date gathered (YYYY-MM-DD) |
source_skill | Yes | Which skill produced this (literature_research, paper_analysis, review_analysis, chapter_outline, or manual) |
chapter | No | Target chapter (e.g., chapter_01) |
tags | Yes | List of topic tags for retrieval |
Filename convention: Descriptive, topic-based, lowercase with underscores. Examples:
nfw_profile_definition.md
gce_interpretations_summary.md
paper4_domain_adaptation_approach.md
Mode 1: Save
Trigger: User says /knowledge save, "save this to knowledge", or at the end of a research skill workflow.
Steps
-
Identify insights: Review the current conversation for key findings worth persisting — definitions, methodology summaries, reference collections, structural decisions, or any reusable knowledge.
-
Check for duplicates: Scan .agent/knowledge/ filenames. If a file on the same topic exists, update it rather than creating a duplicate.
-
Write files: For each insight, create a .md file in .agent/knowledge/ following the format above.
-
Confirm: Report to the user what was saved and the filenames used.
What to Save
- Definitions and standard physics (equations, parameters)
- Methodology summaries from author papers
- Key results and numerical values
- Reference collections for specific topics (with bib keys)
- Structural decisions (how experts organize arguments)
- Transition logic between chapters/sections
What NOT to Save
- Raw NotebookLM responses (too verbose, not curated)
- Conversation-specific context that won't generalize
- Duplicate information already in
references.md files
Mode 2: Retrieve
Trigger: User says /knowledge, "check knowledge for X", or automatically at the start of chapter_outline (Step 0).
Steps
-
List knowledge files: find_by_name(SearchDirectory=".agent/knowledge/", Pattern="*.md")
-
Match by relevance: Compare filenames and read frontmatter to find files matching the current topic. Use these signals:
- Filename keywords: Does the filename contain relevant terms?
- Tags: Do the frontmatter tags overlap with the current topic?
- Chapter: Is the file tagged for the chapter being worked on?
-
Read matched files: Read the body of relevant files (typically 3–8 matches max).
-
Synthesize: Present a summary of what was found, organized by relevance. Highlight:
- Insights that can be used directly
- References already collected (with bib keys)
- Gaps — topics the user is asking about that have NO existing knowledge
Retrieval Strategy
Files found: 12
├── Filename match: 3 files (contain topic keywords)
├── Tag match: 5 files (overlapping tags)
├── Chapter match: 2 files (same chapter)
└── No match: 2 files (skip)
→ Read top 8 matched files
→ Synthesize into context summary
Integration with Other Skills
This skill is called BY other skills:
| Calling Skill | Mode | When |
|---|
literature_research | Save | After producing references.md, save key insights |
paper_analysis | Save | After extracting methodology/results from author papers |
review_analysis | Save | After extracting content from review articles |
chapter_outline | Retrieve | Step 0, before starting research — check for existing knowledge |
Manual (/knowledge) | Either | User explicitly asks to save or retrieve |
Common Mistakes
- Saving too much: Knowledge files should be curated summaries, not raw dumps. If it's longer than ~50 lines, it's probably too detailed.
- Forgetting bib keys: Always include bib keys when referencing papers. Run the lookup against
bibliography.bib.
- Not checking for duplicates: Always scan existing files before creating new ones. Update > duplicate.
- Vague tags: Tags should be specific enough to distinguish topics. Use
[nfw, density-profile] not just [physics].