engram-skill-pack

engramx's AST code-graph + bi-temporal mistakes, as auto-discoverable Claude Code Skills.

v0.2.0 "Structural Bootstrap" — three active skills, two queued for v0.3.0. The pack covers the full "before-you-make-the-edit" loop: orient the agent in a new repo, answer structural questions, surface past mistakes. All under 5 KB of context.

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

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

• engram — core engramx CLI + MCP server + graph store
• engramx-continue — Continue.dev wrapper
• nickcirv.engram-vscode — VS Code / Cursor extension

License

Apache-2.0. Same as engramx.