Memory

Team memory captures accumulated knowledge over time. Unlike skills (which are prescriptive patterns), memories are descriptive facts the team has learned — decisions made, conventions adopted, incidents resolved, and domain knowledge documented.

The @arvoretech/memory-mcp provides semantic search via MCP tools, so AI agents can search your team’s knowledge base on-demand during conversations.

Quick Start

Add the memory section to your hub.yaml:

memory:
  path: ./memories
  categories:
    - decisions
    - conventions
    - incidents
    - domain
    - gotchas
  auto_capture: true

Create your first memory:

hub memory add decisions "Use PostgreSQL for all services" \
  --content "We chose PostgreSQL over MongoDB because we need ACID transactions and complex joins." \
  --tags "database,architecture" \
  --author joao.barros

Run hub generate to inject memories into your editor config.

Categories

Memory File Format

Memories are markdown files with YAML frontmatter, stored in memories/<category>/:

---
title: Use PostgreSQL for all services
category: decisions
date: 2024-06-01
author: joao.barros
tags: [database, architecture]
status: active
---

## Context
We needed to choose between PostgreSQL and MongoDB for our main database.

## Decision
PostgreSQL, because we need ACID transactions and complex joins across entities.

## Consequences
- Migrations managed by Ecto (Elixir) and Prisma (NestJS)
- No dynamic schema flexibility
- Need to manage connection pools carefully

Status Values

• active — Included in searches and prompt injection (default)
• superseded — Replaced by a newer decision, kept for history
• archived — Soft-deleted, excluded from active searches

Semantic Search via MCP

The @arvoretech/memory-mcp server provides semantic search using local embeddings. Add it to your hub.yaml:

mcps:
  - name: team-memory
    package: "@arvoretech/memory-mcp"
    env:
      MEMORY_PATH: ./memories

This gives AI agents access to these tools during conversations:

The MCP uses the paraphrase-multilingual-MiniLM-L12-v2 model by default, supporting Portuguese and English queries. Vectors are stored in a local LanceDB database (memories/.lancedb/) with cosine similarity search and metadata filtering.

Custom Model

Override the embedding model via environment variable:

mcps:
  - name: team-memory
    package: "@arvoretech/memory-mcp"
    env:
      MEMORY_PATH: ./memories
      MEMORY_EMBEDDING_MODEL: Xenova/all-MiniLM-L6-v2  # Smaller, English-only

Auto-Capture

When auto_capture: true, the orchestrator is instructed to extract learnings from completed tasks and save them as new memories. After each task delivery, the AI may create entries for:

• Architectural decisions made during the task
• New conventions discovered
• Bugs found and their root causes
• Domain knowledge clarified

CLI Commands

hub memory list                           # List all active memories
hub memory list --category decisions      # Filter by category
hub memory list --status archived         # Show archived memories
hub memory add <category> "<title>"       # Create a memory
hub memory archive <id>                   # Soft-delete
hub memory remove <id>                    # Permanent delete

Configuration Reference