AI Agentsai-agentsai-toolscodebase-intelligencedeveloper-toolslocal-first
Installation
# Add to your Claude Code skills
git clone https://github.com/pi22by7/In-Memoria
README.md
In Memoria
Giving AI coding assistants a memory that actually persists.
Quick Demo
Watch In Memoria in action: learning a codebase, providing instant context, and routing features to files.
The Problem: Session Amnesia
You know the drill. You fire up Claude, Copilot, or Cursor to help with your codebase. You explain your architecture. You describe your patterns. You outline your conventions. The AI gets it, helps you out, and everything's great.
Then you close the window.
Next session? Complete amnesia. You're explaining the same architectural decisions again. The same naming conventions. The same "no, we don't use classes here, we use functional composition" for the fifteenth time.
Every AI coding session starts from scratch.
This isn't just annoying, it's inefficient. These tools re-analyze your codebase on every interaction, burning tokens and time. They give generic suggestions that don't match your style. They have no memory of what worked last time, what you rejected, or why.
The Solution: Persistent Intelligence
In Memoria is an MCP server that learns from your actual codebase and remembers across sessions. It builds persistent intelligence about your code (patterns, architecture, conventions, decisions) that AI assistants can query through the Model Context Protocol.
Learns your patterns - Analyzes your code to understand naming conventions, architectural choices, and structural preferences
Instant project context - Provides tech stack, entry points, and architecture in <200 tokens (no re-analysis needed)
Smart file routing - Routes vague requests like "add password reset" directly to relevant files
Semantic search - Finds code by meaning, not just keywords
Work memory - Tracks current tasks and architectural decisions across sessions
Pattern prediction - Suggests how you'd solve similar problems based on your history
Example Workflow
# First time: Learn your codebase
npx in-memoria learn ./my-project
# Start the MCP server
npx in-memoria server
# Now in Claude/Copilot:
You: "Add password reset functionality"
AI: *queries In Memoria*
"Based on your auth patterns in src/auth/login.ts, I'll use your
established JWT middleware pattern and follow your Result<T>
error handling convention..."
# Next session (days later):
You: "Where did we put the password reset code?"
AI: *queries In Memoria*
"In src/auth/password-reset.ts, following the pattern we
established in our last session..."
No re-explaining. No generic suggestions. Just continuous, context-aware assistance.
Quick Start
Installation
# Install globally
npm install -g in-memoria
# Or use directly with npx
npx in-memoria --help
Connect to Your AI Tool
Claude Desktop - Add to your config (~/Library/Application Support/Claude/claude_desktop_config.json):
# Analyze and learn from your project
npx in-memoria learn ./my-project
# Or let AI agents trigger learning automatically
# (Just start the server and let auto_learn_if_needed handle it)
npx in-memoria server
How It Works
In Memoria is built on Rust + TypeScript, using the Model Context Protocol to connect AI tools to persistent codebase intelligence.
Pattern learner (statistical analysis of your coding style)
Semantic engine (understands code relationships and concepts)
TypeScript Layer - MCP server and orchestration:
13 specialized tools for AI assistants (organized into 4 categories)
SQLite for structured data, SurrealDB with SurrealKV for persistent vector embeddings
File watching for incremental updates
Smart routing that maps features to files
Storage - Local-first:
Everything stays on your machine
SQLite for patterns and metadata
SurrealDB with SurrealKV backend for persistent vector embeddings
Local transformers.js for embeddings (Xenova/all-MiniLM-L6-v2)
What Makes It Different
This isn't just another RAG system or static rules engine:
Learns from actual code - Not manually-defined rules, but statistical patterns from your real codebase
Predicts your approach - Based on how you've solved similar problems before
Token efficient - Responses optimized to minimize LLM context usage (<200 tokens for project context)
Routes to files - "Add login" → automatically suggests src/auth/login.ts
Remembers context - Tracks work sessions, tasks, and architectural decisions
Multi-mode search - Semantic (meaning), text (keywords), or pattern-based
What's New in v0.5.x
We recently completed Phases 1-4 of the implementation roadmap:
🗺️ Project Blueprints (Phase 1)
Instant project context without full learning. Ask about a codebase and get tech stack, entry points, key directories, and architecture all in under 200 tokens.
💼 Work Context System (Phase 2)
AI agents can now track work sessions, maintain task lists, and record architectural decisions. Resume work exactly where you left off.
🧭 Smart File Routing (Phase 3)
Feature-to-file mapping across 10 categories (auth, API, database, UI, etc.). Vague requests like "add password reset" get routed to specific files automatically.
⚡ Smooth Progress Tracking (v0.5.3)
No more janky console spam. Progress bars update in-place with consistent 500ms refresh rates.
MCP Tools for AI Assistants
In Memoria provides 13 specialized tools that AI assistants can call via MCP. They're organized into 4 categories (down from 16 after Phase 4 consolidation merged redundant tools):
🎯 Core Analysis (2 tools)
analyze_codebase - Analyze files/directories with concepts, patterns, complexity (Phase 4: now handles both files and directories)
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
