Skip to content

NickCirv/engram-skill-pack

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

engram-skill-pack

One command installs engram in Claude Code. Three auto-discoverable Skills that catch recurring mistakes, answer structural questions, and bootstrap repo understanding. Wraps engramx — the underlying AST + bi-temporal mistakes engine — into the Anthropic Claude Code Marketplace format.

# Install both at once:
npm i -g engramx && npx @anthropic-ai/claude-code add nickcirv/engram-skill-pack

v0.2.0 "Structural Bootstrap" ships three active skills. Two more queued for v0.3.0 (gen + learn). Full "before-you-make-the-edit" loop in under 5 KB of context.

License: Apache 2.0 npm requires engramx skills tests marketplace bundle cloud node

Skills: engram-mistakes · engram-query · engram-gods · engram-gen (v0.3) · engram-learn (v0.3)

Sections: Install · The Five Skills · How is this different from claude-mem? · FAQ · Roadmap · Privacy


See it in 30 seconds

engram v4.0 demo — bi-temporal mistakes auto-captured from git revert (30 seconds)

The pack ships these skills as the Claude Code Marketplace surface for what the recording shows. Same engramx underneath, different discovery mechanism.


Three questions answered in 30 seconds

Q: Why install this if I already have engramx? The skill-pack is the Claude Code Marketplace surface for engramx. Same data, different discovery path — Claude Code's auto-discovery mechanism fires the right engramx query when you ask "any mistakes here?" or "what's important in this repo?" without you having to remember the CLI. Install it alongside engramx, not instead of.

Q: What if I use Cursor / Continue / VS Code instead of Claude Code? You don't need this pack. Different IDEs have different engramx wrappers — engramx-continue (npm) for Continue.dev, nickcirv.engram-vscode (OpenVSX) for VS Code/Cursor. All three plus this pack read the same ~/.engram/graph.db underneath. Install engramx once, surface it everywhere.

Q: Does this leak my code anywhere? No. The pack ships zero egress. engramx (the peer dependency) routes all graph data through local SQLite at ~/.engram/ and <project>/.engram/. No network calls during install or skill invocation. Safe on air-gapped workstations. Apache-2.0.


What it does

When your AI agent is about to edit a file, engram-mistakes auto-fires and injects bi-temporal context — what was once believed about the code, when that belief was found false, and what the truth is now. The agent reads the warning before making the edit and adapts.

⛔ engramx pre-mortem — you've made this mistake before:

  ⚠ useReducer + async + form-event handlers
     ┌─ then you believed: useReducer dispatch is safe in onChange handlers
     ├─ found false:       2026-04-19
     └─ truth now:         useReducer with async dispatch needs useCallback wrapping
        file: src/forms/ContactForm.tsx

The skill doesn't invent new primitives. It wraps the existing engramx CLI with the auto-discovery metadata the Claude Code skills marketplace expects.

Install

# Install the pack
npm i -g engram-skill-pack

# Add to your Claude Code via marketplace
npx @anthropic-ai/claude-code add nickcirv/engram-skill-pack

Requires engramx >= 4.0.0 installed and engram init run in any project you want context for.

# One-time engramx install (if you don't have it)
npm i -g engramx

# In each project where you want bi-temporal mistakes
cd ~/your-project && engram init

The init command auto-installs the Sentinel hook into the project's .claude/settings.local.json. The hook fires PreToolUse and is permissive by default. Opt out via engram init --no-hook or ENGRAM_MISTAKE_GUARD=0.

The five skills

Skill Status What it does
engram-mistakes ✅ v0.1.0 Surface bi-temporal mistakes (the rave moment — auto-fires PreToolUse)
engram-query ✅ v0.2.0 Natural-language structural queries — replaces ad-hoc Grep
engram-gods ✅ v0.2.0 Bootstrap repo understanding — top-N entities by importance
engram-gen 🚧 v0.3.0 Generate task-grounded context prompts
engram-learn 🚧 v0.3.0 Capture bi-temporal mistakes manually (interactive 4-field prompt)

How is this different from claude-mem?

engram-skill-pack claude-mem serena hermes caveman
Memory primitive Structural (AST + git graph) Functional (literal context capture) Functional Functional Functional
Bi-temporal mistakes Yes — then-believed vs truth-now No No No No
Auto-capture from git revert Yes (v4.0+) No No No No
PreToolUse hook Yes — fires before edit, surfaces past corrections No No No No
Local-first / zero cloud Yes Yes Yes Mixed Yes
Multi-IDE support Yes — Claude Code, Continue, VS Code, Cursor Claude Code only Editor extension Various Cursor focus

The structural moat is the moat. claude-mem captures what was said in a session. engramx captures what's structurally true about the codebase — and what your team learned was false. That's what catches your agent before it repeats a fix the repo already burned to revert.

FAQ

Q: Do I need engramx installed? Yes. The pack is a thin Skills-marketplace wrapper around the engramx CLI. Install once: npm i -g engramx. The pack's postinstall script detects the binary and prints a friendly fix if it's missing — install never fails.

Q: Does this work with Cursor / Continue.dev / VS Code? The pack itself ships only the Claude Code Skills surface. For other IDEs, install the matching wrapper that reads the same ~/.engram/graph.db:

  • Continue.dev → npm i -g engramx-continue
  • VS Code / Cursor → nickcirv.engram-vscode on OpenVSX

All four wrappers (this pack + the three above) read the same underlying engramx graph. Install once, surface everywhere.

Q: How is engram-mistakes different from claude-mem? claude-mem captures whatever you said in a session — literal context strings that the agent can re-read later. engram-mistakes captures what was structurally true about your code (AST entities, files, patterns) and what your team learned was false (git reverts, session corrections, explicit engram learn captures). The bi-temporal axis is the wedge: claude-mem can't tell you what you used to believe versus what you now know.

Q: What's the AAA gate I see referenced? engramx skills follow an internal "AAA" template (control dials + decision tree + variants + quality gate) — the same shape your existing ~/.claude/skills/_aaa-template/ defines. For the marketplace surface, the gate is loosened to: valid frontmatter, name matches manifest, description ≤250 chars, SKILL.md exists. The full AAA bar is enforced for engram-internal skills, not Anthropic Marketplace skills.

Q: Privacy? Zero egress in this pack. engramx (peer dependency) routes all graph data through local SQLite at ~/.engram/ and <project>/.engram/. No network calls happen during install or skill invocation. Safe on air-gapped workstations.

Roadmap

  • v0.1.0 (2026-05-18) — Memory Preview: engram-mistakes only ✅
  • v0.2.0 (2026-05-18) — Structural Bootstrap: adds engram-query + engram-gods
  • v0.3.0 (~7 days) — Capture Layer: adds engram-gen + engram-learn with the interactive 4-field bi-temporal prompt
  • v0.4.0 — AAA-elevation pass (full Layer 1-7 across all 5 skills + cross-ecosystem sync)
  • v1.0.0 — Marketplace launch, comparison-benchmark page (engramx-bench.dev) live, Show HN

Related repos

License

Apache-2.0. Same as engramx.

About

engramx's bi-temporal mistakes + structural code-graph as auto-discoverable Claude Code Skills. Catches recurring mistakes before agents repeat them. Apache 2.0, local SQLite, zero cloud.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors