engram

Engram

A memory system for Claude Code that learns how you work and generates new skills from your patterns.

Installation

As a Claude Code Plugin (Recommended)

Local Development:

# Clone the repo
git clone https://github.com/chietala/engram.git
cd engram
bun install

# Run Claude Code with the plugin
claude --plugin-dir /path/to/engram

From GitHub:

# Add the marketplace
/plugin marketplace add chietala/engram

# Install the plugin
/plugin install engram@engram-marketplace

Team Auto-Install: Add to your project's .claude/settings.json:

{
  "extraKnownMarketplaces": {
    "engram-marketplace": {
      "source": {
        "source": "github",
        "repo": "chietala/engram"
      }
    }
  },
  "enabledPlugins": {
    "engram@engram-marketplace": true
  }
}

The Idea

Every time you use Claude Code, you're teaching it something. Engram captures those lessons automatically:

• Watch: Observes your tool usage (file reads, bash commands, edits)
• Remember: Stores patterns with semantic understanding
• Evolve: Challenges and refines patterns through contradiction
• Generate: Creates reusable Claude Skills from mature patterns

Plugin Commands

Once installed, these commands are available:

Command	Description
/engram-status	Show memory counts and system status
/engram-query <search>	Search memories semantically
/engram-generate <name>	Generate a skill from recent patterns
/engram-publish	Publish mature patterns as Claude rules

Features

Automatic Pattern Detection

Engram notices when you do similar things repeatedly. Read a config file then run a command? Do it a few times and Engram recognizes the pattern.

Hegelian Learning

Patterns evolve through dialectic:

1. Thesis - Initial pattern: "Always use git status before committing"
2. Antithesis - Contradiction: "But sometimes you skip it for quick fixes"
3. Synthesis - Resolution: "Use git status for multi-file changes, skip for single-line fixes"

This creates nuanced, context-aware knowledge.

LLM-Powered Insights

Claude analyzes your patterns and generates insights like:

"The user is building a self-improving system that observes its own tool usage patterns and automatically generates new reusable skills"

Native Memory Integration

When patterns reach synthesis, Engram automatically publishes them as Claude rules—meaning Claude remembers your patterns without the plugin running:

.claude/rules/engram/
├── testing-workflow.md      # "Run tests before committing"
├── error-handling.md        # "Check logs after deployments"
└── code-review.md           # "Read tests before reviewing PRs"

Rules include:

• Path-specific triggers: Only activate for relevant file types
• Linked skills: Reference executable skills for automation
• Confidence tracking: Update as patterns evolve

Example generated rule:

---
paths:
  - "**/*.ts"
  - "src/**"
---

# Test Before Commit

Before committing any code changes, always run the test suite with `bun test`.

## Related Skill

This pattern has an associated skill: **test-before-commit**

> Runs tests and commits changes with conventional commit messages

**To invoke this skill**, say:
- "commit my changes"
- "run tests and commit"
- "safe commit"

Automatic Skill Generation

When a pattern matures enough (used consistently, survived contradictions), Engram generates a Claude Skill you can use in future sessions.

Hybrid Skill Scripts

Generated skills aren't just static instructions—they're executable scripts that combine:

• Deterministic code for file operations, glob patterns, shell commands
• Intelligence points for LLM-powered reasoning where judgment is needed

// Deterministic: gather data
const files = await glob.scan({ pattern: "**/*.ts" });
const diff = await $`git diff --cached`.text();

// Intelligence point: generate insights
const summary = await intelligence(
  "Summarize the key patterns in this codebase",
  { files: fileList, diff }
);

// Deterministic: output results
console.log(summary);

This creates scripts that are both reliable and intelligent.

Memory Consolidation

Like human memory:

• Working memory - Current session activity
• Short-term memory - Recent patterns, subject to decay
• Long-term memory - Proven patterns that persist

Old, unused memories fade. Important ones strengthen.

Semantic Search

Query your memories naturally:

/engram-query "how do I handle errors"

Finds relevant patterns even if wording differs.

What Gets Captured

Event	What's Learned
File reads	What you look at to understand code
Bash commands	Your debugging and build workflows
File edits	How you modify code
Prompts	What you're trying to accomplish