MCP Protocol · Azure Native · GitHub

Give your AI agents memory.

Persistent, searchable, three-tier memory for AI coding agents. Store decisions, learn from patterns, and share knowledge across sessions — automatically.

Quick Start Request Access

MCP Tools

Memory Tiers

1536

Embedding Dims

0ms

Proxy Overhead

The Problem

Every session starts from zero.

Your AI agents are brilliant — but they have amnesia. Every conversation, every decision, every pattern learned, gone the moment the session ends.

Context Amnesia

Agents forget decisions, architecture choices, and rationale. You re-explain the same context every single session.

Repeated Mistakes

Without memory, agents make the same mistakes and propose solutions that were already rejected in prior sessions.

Knowledge Silos

When team members leave or switch projects, institutional knowledge disappears. Onboarding starts from scratch.

Architecture

Three tiers of memory.

Information naturally flows from ephemeral conversations to permanent organizational knowledge — just like human memory.

Short-Term

Conversations

TTL: 7 days · /threadId

Ephemeral context within active conversations. Fast, disposable, high-volume.

Current coding session turns
File read/write context
Tool usage captures

Medium-Term

Project Memory

TTL: 90 days · /projectId

Decisions, summaries, and artifacts scoped to a project. The working memory of your team.

“We chose PostgreSQL for ACID”
Architecture decision records
Session summaries & artifacts

Long-Term

Knowledge Base

Permanent · /tenantId

Validated patterns and best practices promoted to org-wide knowledge. Permanent, multi-tenant.

API error handling pattern
Microservices gateway standard
Deployment checklist template

Capabilities

Everything agents need to remember.

Semantic Vector Search

Find relevant memories by meaning, not just keywords. 1536-dimension embeddings powered by Azure OpenAI with hybrid search combining 70% vector + 30% keyword scoring.

bmad_hybrid_search · bmad_vector_search

Knowledge Graph

Bidirectional links between memories — supersedes, implements, rationale, related-to. Graph traversal up to 3 hops with cycle detection and auto-surfacing.

bmad_link_memories · bmad_get_linked

Session Tracking

Automatic capture of every file read, edit, and command via Claude Code hooks. Local-first JSONL logs with optional git branch-per-session isolation.

capture.js hooks · session-manager.js

Conversation Proxy

Transparent API proxy captures every prompt and response between Claude Code and Anthropic. Zero latency overhead, automatic secret redaction, two-file JSONL logging.

proxy/ · conversations.jsonl · turns.jsonl

Role-Based Behavior

Architects get high proactivity and frequent storage. Developers get minimal interruption. Five integration levels from menu-only to full automation.

5 integration levels · 14+ pre-built agents

Graceful Degradation

Three-state fallback: full MCP memory → local-only session tracking → session-only context. Never breaks the agent. Works offline, syncs when available.

offline-first · 3-state model

Data Flow

From keystrokes to knowledge.

Every interaction flows through a capture pipeline — from real-time conversation proxying to permanent organizational knowledge.

Claude Code

Agent + User

→

Proxy

Capture + Redact

→

Anthropic API

LLM Processing

↓ async, non-blocking

JSONL Logs

turns + conversations

Cosmos DB

3-tier memory

Session JSONL

hooks capture

User-facing

Transparent proxy

Local storage

Cloud storage

Getting Started

Three steps to agents that remember.

Install

Add the MCP server to your project. Deploy Cosmos DB with the included ARM template or connect to an existing instance.

npm install @acme-corp/agent-memory-server

Configure

Point Claude Code at the MCP server. Add your Azure Cosmos DB and OpenAI credentials. Optionally start the conversation proxy.

// .claude/settings.local.json
"mcpServers": { "bmad-memory": { ... } }

Remember

Your agents now have memory. Decisions persist. Patterns compound. Knowledge promotes from project to organization — automatically.

bmad_hybrid_search("auth pattern")
→ 3 relevant memories found

Quick Start

Up and running in 5 minutes.

.claude/settings.local.json

# 1. Install the MCP server
npm install @acme-corp/agent-memory-server

# 2. Deploy Azure infrastructure (optional — dev template included)
az deployment group create \
  --resource-group bmad-memory-dev-rg \
  --template-file infra/azuredeploy.json \
  --parameters infra/azuredeploy.parameters.dev.json

# 3. Configure Claude Code MCP server
{
  "mcpServers": {
    "bmad-memory": {
      "command": "node",
      "args": ["node_modules/@acme-corp/agent-memory-server/dist/index.js"],
      "env": {
        "COSMOS_ENDPOINT": "https://your-account.documents.azure.com:443/",
        "COSMOS_KEY": "your-cosmos-key",
        "AZURE_OPENAI_ENDPOINT": "https://your-openai.openai.azure.com/",
        "AZURE_OPENAI_API_KEY": "your-openai-key"
      }
    }
  }
}

# 4. (Optional) Start the conversation proxy
cd proxy && npm start
$env:ANTHROPIC_BASE_URL="http://localhost:8642"; claude