@shahmilsaari/memory-core

v1.0.53

Published

6 hours ago

Universal AI memory core — generate AI context files from architecture profiles with RAG support

0High
0Medium
0Low

shahmilsaari

ai memory rag architecture copilot cursor claude context

memory-core

Universal AI memory for developers. Store architecture rules once — every AI coding agent reads them before writing code.

The problem

AI tools like Copilot, Cursor, and Claude Code start fresh every session. They don't know your architecture decisions, your layer boundaries, or your team conventions. They write code that violates patterns you spent months establishing.

memory-core fixes that. Store your rules once, enforce them everywhere.

What it does

Remember — store architecture decisions and rules in PostgreSQL with plain-English reasons
Distribute — generate instruction files for 14 AI agents automatically from your stored rules
Enforce — multi-tier pipeline: deterministic layer graph (instant, confidence 1.0) → AI semantic review → optional model critique second opinion → confidence gate arbitrates. Auto-fixes violations on save by default.

npx @shahmilsaari/memory-core init

Key features

281 pre-built rules — seed best-practice rules for your stack in one command, with plain-English reasons
14 AI agents supported — generates instruction files for Copilot, Cursor, Claude Code, Windsurf, and more automatically
Evidence-based enforcement — deterministic AST + dependency graph catches layer violations instantly; AI semantic review with confidence gating catches the rest; optional AI auto-fix on save
Runs fully local — embeddings stay on your machine via Ollama; use any cloud provider (DeepSeek, OpenAI, Anthropic) only for code analysis
Live dashboard — real-time violations feed, interactive architecture graph, memory model-check with batch resolve, and model switcher at localhost:5178
Team-ready — export rules to memories.json, commit it, and teammates import with one command — no shared database needed
Token-optimized — compressed prompts, prompt caching, and confidence-gated model critique save ~58% tokens per check without sacrificing accuracy
CI/CD built-in — generates a GitHub Actions workflow; CI reads memories.json directly, no Ollama required
Commit message linting — enforce conventional commits or any regex pattern via the commit-msg hook
Violation analytics — track which rules fire most and auto-tune noisy ones with memory-core tune
Pre-commit hook — advisory or strict mode; bypass anytime with MEMORY_CORE_SKIP_HOOK=1

Get started

Requires: Node.js ≥ 18, PostgreSQL 14+, Ollama with nomic-embed-text

# 1. Initialize in your project
npx @shahmilsaari/memory-core init

# 2. Load best-practice rules for your stack
memory-core seed --arch clean-architecture

# 3. Store your own decisions
memory-core remember "No direct DB calls from controllers"

# 4. Enforce on every commit
memory-core hook install

That's it. Every AI agent in your project now reads your rules before writing code.

Supported agents

Claude Code · GitHub Copilot · Cursor · Windsurf · Cline · Roo Code · Aider · Continue.dev · Devin · Amazon Q · Gemini Code Assist · Zed AI · JetBrains AI · OpenHands

Architecture profiles

281 pre-built rules ready to load:

clean-architecture · nestjs · react · svelte · vue · hexagonal · modular-monolith · mvc · laravel-service-repository · nuxt · angular · go-api · react-native

Commands

memory-core remember "..."          # store a rule or decision
memory-core sync                    # push rules to all AI agent files
memory-core check --diff HEAD~1     # enforce architecture on a diff (deterministic by default)
memory-core check --diff HEAD~1 --ai # same, but with AI analysis
memory-core watch                   # real-time violation detection on save
memory-core watch --auto-fix        # also auto-rewrite violating files via AI
memory-core dashboard               # live command center at localhost:5178
memory-core export                  # share rules with your team via memories.json
memory-core stats                   # see which rules fire most

→ Full command reference

Environment

Local Ollama (default):

DATABASE_URL=postgresql://localhost:5432/memory_core

EMBEDDING_URL=http://localhost:11434
EMBEDDING_MODEL=nomic-embed-text
EMBEDDING_TIMEOUT_MS=30000

CHAT_PROVIDER=ollama
CHAT_MODEL=qwen2.5-coder:7b
CHAT_URL=http://localhost:11434
CHAT_TIMEOUT_MS=120000

Remote Ollama (different host for embedding and chat):

DATABASE_URL=postgresql://localhost:5432/memory_core

EMBEDDING_URL=http://192.168.1.10:11434
EMBEDDING_MODEL=nomic-embed-text
EMBEDDING_TIMEOUT_MS=30000

CHAT_PROVIDER=ollama
CHAT_MODEL=qwen2.5-coder:7b
CHAT_URL=http://192.168.1.10:11434
CHAT_TIMEOUT_MS=120000

Cloud provider for chat, local Ollama for embedding:

DATABASE_URL=postgresql://localhost:5432/memory_core

EMBEDDING_URL=http://localhost:11434
EMBEDDING_MODEL=nomic-embed-text

CHAT_PROVIDER=openai
CHAT_MODEL=gpt-4o-mini
CHAT_API_KEY=sk-...

EMBEDDING_URL and CHAT_URL are independent — point them to different hosts if needed. EMBEDDING_TIMEOUT_MS defaults to 5 s; increase to 30 s+ for remote Ollama.

License

MIT