A knowledge management tool that saves session findings and decisions, allowing future sessions to start with full context.

Install

mkdir -p .claude/skills/skills-jackccrawford && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/15420" && unzip -o skill.zip -d .claude/skills/skills-jackccrawford && rm skill.zip

Installs to .claude/skills/skills-jackccrawford

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

Your next session starts cold. No memory of what you built, what broke, what you decided. Every memory you write is a gift to that future session. The richer the memory, the less time re-learning.
196 chars · catalog descriptionno explicit “when” trigger
Beginner

Key capabilities

  • →Save session memories to a SQLite database
  • →Recall memories by meaning, not just keywords
  • →Tag memories with gists for quick identification
  • →Thread follow-up memories to existing ones
  • →Bulk-load existing files as memories
  • →Build and manage a local BERT model for semantic search

How it works

Memories are stored in a SQLite database, and semantic search is performed using a local BERT model that finds memories by meaning rather than keywords.

Inputs & outputs

You give it
text 'Token validation was running before refresh in auth.rs.'
You get back
a saved memory with gist 'fix: auth token refresh , async ordering in middleware'

When to use skills

  • →Save session notes
  • →Recall past debugging findings
  • →Document architectural decisions
  • →Maintain technical history

About this skill

geniuz

Your next session starts cold. No memory of what you built, what broke, what you decided. Every memory you write is a gift to that future session. The richer the memory, the less time re-learning.

Works with any framework — Claude Code, Cursor, Windsurf, Aider, or anything with shell access.

AI tools use the CLI commands below: geniuz keep writes, geniuz recall reads. The terminal UI (geniuz tui) and the dashboard are for people to browse their memories. The TUI refuses to launch from non-interactive callers and exits with code 2, so calling it from a subprocess locks nothing up.

Why this matters

A memory that says "fixed auth bug" saves your future self zero time.

A memory that says what broke, why it broke, and how you fixed it saves an hour.

Memories compound. A single memory is a note. A folder of memories is institutional knowledge. Write for the version of you that knows nothing.

Asking first

Most people like to be asked, once, whether notes should be kept in Geniuz: decisions, progress, what to pick up next time. When they agree, keeping as you go and mentioning what was kept works well. It helps in this chat and in the next. In a scheduled or unattended run, with no one to ask, keeping what the next run will need is what Geniuz is for.

Commands

Every keep names its folder (--in) and its gist (-g). A plain recall reads My memory (or the folder named in .geniuz-folder); add --in Work to read Work, or --all for every folder.

# A fresh install has one folder, "My memory". These examples keep in a folder
# called Work, made once with:
geniuz folder create Work -d "Fixes, decisions and progress on the code."

# Save what you learned — pipe in for depth, inline for quick notes
echo "Token validation was running before refresh in auth.rs.
Swapped lines 42-47. Root cause: middleware ordering assumed
sync validation, but OAuth refresh is async. Three edge cases
tested: expired token, revoked token, concurrent refresh." \
  | geniuz keep --in Work -c - -g "fix: auth token refresh — async ordering in middleware"

# Quick memory when depth isn't needed
geniuz keep --in Work -c "Upgraded rusqlite to 0.32" -g "dep: rusqlite 0.32"

# Keep from a file
geniuz keep --in Work -c @session-notes.md -g "session: March 19 architecture review"

# Thread a follow-up
geniuz keep --in Work -c "Same fix needed in staging compose" -g "fix: staging auth ordering" -p A1B2C3D4

# Search: by meaning and by words, together. There is no mode to choose
geniuz recall --in Work "authentication middleware"
geniuz recall --in Work "what broke in production" --since 7d

# Proper names match exactly and come first (a Capitalized word after the
# first, or anything in "quotes")
geniuz recall "who reviewed the PR, was it Cubic"

# Every line holding exact words, from full content, newest first
geniuz recall --grep Cubic --since 7d
geniuz recall --grep Cubic --grep "pull request"   # every term must appear
geniuz recall --all --grep Cubic                  # the same, in every folder

# A whole thread, root first, from any memory in it
geniuz recall --thread A1B2C3D4

# Recent memories: recall with no search
geniuz recall
geniuz recall -l 5

# Full content (not just gist)
geniuz recall --full "auth"

# Discover something you forgot you knew
geniuz recall --random

# Bulk-load existing files
geniuz capture ./docs/

# Build embedding cache (once, then automatic)
geniuz backfill

# Stats for the folder in use
geniuz status

Folders

Memories live in folders, each with a name and a description. Choose the folder before you keep, because a memory can't be moved afterwards. Read its description first: it may say what belongs there, what must not go there, and how to use it.

Make a folder for anything you'll come back to: a project, a topic, a long piece of work. In a short chat or a long one, keep what matters as you go. Don't count on the chat's history to hold it: a chat gets compacted, closed or forgotten, and a folder keeps what you put in it.

geniuz folder create "Website rebuild" -d "Decisions and progress on the new site."

The description is for whoever opens the folder next. Often that's you. Through MCP, pass create to the folders tool.

A folder can have a face: one emoji shown beside its name in the Geniuz window, chosen by whoever writes there. geniuz folder face "Website rebuild" 🧱 sets it (through MCP, face on the folders tool). Without one, the window shows a small mosaic laid out from the folder's ID.

geniuz folder list                     # all of them; * is the one in use
geniuz folder show Work                # one in full, with its description

geniuz keep --in Work -c "..." -g "..."
geniuz recall --in Work "deploy"
geniuz recall --all "deploy"           # every folder; each result shows its address
geniuz recall 13625F6B:3487B9B0        # open an address, whichever folder it's in

A wrong name is an error, never a different folder. A .geniuz-folder file holding a folder's name or ID decides what a plain geniuz recall reads in that directory (unless GENIUZ_DB is set, which wins); a keep still names its folder.

Multiple agents can use the same folder. What one agent learns, every agent finds. That also makes a folder a way to leave each other messages: thread your reply to theirs with -p.

Never copy a folder's files: a copy of a live database can lose memories, and a folder can't be duplicated, because its ID and its memories' IDs are permanent. To relocate one, use geniuz folder move. To bring one from another computer, export it there (geniuz folder export NAME file.db) and import the file here (geniuz folder import file.db, with --as "Name" if its name is taken): it arrives as the same folder with the same IDs. Nothing merges, and a folder already here is refused, so stop keeping at the source before exporting.

When to keep

  • After solving something hard. Your next session will hit the same problem. Give it the answer.
  • Before your session ends. Everything in your context window dies with the session. Your folder doesn't.
  • When you discover something non-obvious. The small facts — which config flag, which line number, which edge case — are the ones that save the most time.

When to recall

  • Starting a new task. You may have already solved part of it last week.
  • When stuck. The answer might be in a memory from three sessions ago.
  • After compaction. Your context was compressed. Your folder wasn't.

How to find things

A search is for finding: it ranks memories by what they mean and by the words they hold, so a topic, a name or a term all work in one query. --grep is for listing: the lines that hold a word, newest first, up to the limit (50 lines unless you raise it with -l).

  1. If the question has a time in it ("last week", "yesterday"), add --since: 7d, 24h, 2w, or a date. Without it, months of older memories outrank last week's answer.
  2. Start with a search. Every answer ends with names seen: and pointers:. Those are your next hops.
  3. List exactly what you found: --grep Devin, or --thread 3487B9B0 to read the whole conversation.
  4. Read the last line. searched: 312 memories · since 7d · grep "Cubic" says what "nothing found" means: nothing in that scope.

Through MCP the same works as recall arguments: folder, all, since, until, grep (a word or a list), thread. query takes an address too.

Note: geniuz recall help prints recall's help instead of searching for the literal word "help". To find memories about that word, use a phrase containing it — e.g. geniuz recall "help with debugging" or geniuz recall "how to help".

Writing good memories

The gist is how your future self finds this memory. The content is what makes it useful when found.

Gist: compress the insight. "fix: auth token refresh — async ordering in middleware" — category, what, why, where.

Content: make it self-contained. If your future self reads only this memory — no session history, no surrounding context — can they understand what happened and act on it?

A memory doesn't need to be long. It needs to be complete.

Anything longer than a sentence: use a file (-c @note.md) or stdin (-c -). The inline form is the one a shell quietly corrupts — backticks get evaluated, quotes get eaten, newlines collapse — and you cannot take back what was kept. From Windows PowerShell 5.1, always use one of these two: it passes text with double quotes to geniuz split apart. Prefer the file for anything beyond plain ASCII, since PowerShell 5.1 can pipe accents and dashes as ?.

When something works, write down what the green light did not tell you. A passing test proves something narrower than you meant to ask. That gap is the difference between a note and a lesson.

Images

A memory can show up to five images: a screenshot, a chart, a photo of a whiteboard.

# Keep a memory with an image (repeat --image for more)
geniuz keep -f Work -c "The error dialog, before the fix" -g "bug: save dialog" --image shot.png

# Show an image that is already kept, in another memory
geniuz keep -f Work -c "Same dialog after the fix" -g "fix: save dialog" --image-id 1B2C3D4E

# The images in a folder, and one image's details
geniuz image -f Work
geniuz image -f Work 1B2C3D4E

# Take one out to look at it
geniuz image -f Work 1B2C3D4E -o shot.webp

A result line names a memory's images after the gist (· image 1B2C3D4E). The picture itself is only loaded when you ask for it by that ID, so recalling a memory costs nothing extra until you look.

The words are what search finds, so say in the content what the image shows and why it was kept. Those words stay for good, whatever later happens to the image. You can decline to keep an image; say so plainly if you do.

When you look at an image, or at anything else you recall, the compa


Content truncated.

When not to use it

  • →When the user needs a real-time, in-context memory that dies with the session
  • →When the task does not involve saving or recalling past findings
  • →When the user prefers a graphical dashboard over CLI commands for memory management

Limitations

  • →The terminal UI refuses to launch from non-interactive callers
  • →The BERT model downloads once (~118MB) on first search
  • →GENIUZ_STATION environment variable is used for shared folders

How it compares

This skill provides a persistent, semantic memory system that operates offline and across sessions, allowing recall by meaning, unlike temporary context windows or keyword-only search.

Compared to similar skills

skills side by side with the closest alternatives in the catalog.

SkillInstallsUpdatedSafetyDifficulty
skills (this skill)06moReviewBeginner
knowledge-ops07moNo flagsAdvanced
knowledge-base19moReviewBeginner
notion-knowledge-capture101yNo flagsIntermediate

Try saying

Example prompts that trigger this skill in your AI assistant.

Search skills

Search the agent skills registry