🧠 AI Memory Layer

The AI Memory Layer is a production-ready "Postgres for AI agent memory." It provides a persistent, semantic, and secure infrastructure that gives any AI coding assistant (Cursor, Claude Code, Copilot) true long-term memory about your project, architecture, and organizational decisions.

---

🚀 Enterprise Features (v1.2)

• 🧠 AST-Aware Ingestion: Uses Tree-Sitter to parse code into structural context, capturing function signatures and architectural patterns rather than just text diffs.
• ⚡ HNSW Vector Indexing: Sub-millisecond vector search at scale using Hierarchical Navigable Small World indexes in pgvector.
• 🏎️ Semantic Caching: Redis-backed caching layer that intelligently stores and retrieves common queries to reduce LLM latency and costs.
• 🔭 Full Observability: OpenTelemetry instrumentation for FastAPI and LLM calls, providing deep tracing into retrieval scores and cost monitoring.
• 🛡️ Memory Consolidation: A background self-healing process that distills redundant micro-memories into high-level architectural insights.
• 🤖 CI/CD Integration: Drop-in GitHub Action templates for automated project brain synchronization on every push.

---

💡 Why AI Memory Layer?

Generic chat logs aren't enough for complex engineering. AI Memory Layer is purpose-built for high-stakes software development:

• Zero Lock-In: Run entirely locally using sentence-transformers and Ollama, or scale with OpenAI/Anthropic.
• Architectural Intelligence: We don't just store chat logs. We ingest Git history, auto-detect conflicts, and extract structured taxonomy (episodic, semantic, procedural).
• Enterprise Security: Built-in Multi-Tenancy (project_id) and X-API-Key authentication.
• True Hybrid Search: Combines keyword precision with semantic depth, weighted by recency.

---

🏗️ Architecture

graph TD
    A[Git / Conversations] -->|Ingest + Dedupe Hash| B(Ingestion Pipeline)
    B -->|Summarize & Detect Conflicts| C{LLM: Local/OpenAI/Anthropic}
    C -->|Vector + Keyword + Metadata| D[(Postgres + pgvector + tsvector)]
    E[Agent: Claude/Cursor] -->|MCP / REST| F(Retrieval Engine)
    F -->|Hybrid Search: BM25 + Vector + Recency| D
    D -->|Ranked Memories| F
    F -->|Contextual Response| E

Loading

---

🛠️ Quick Start

1. Spin up Infrastructure

docker-compose up -d

2. Configure Environment

cp .env.example .env
# Edit .env to set your LLM_PROVIDER and API keys

3. Install & Run

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

# Start the API server
uvicorn src.main:app --reload

---

🔌 Agent Integration (MCP)

Give your AI agent a "long-term brain" by connecting it to the Model Context Protocol (MCP) server.

Cursor / Windsurf

1. Go to Settings -> Models -> MCP Servers.
2. Add a new server:
   • Type: command
   • Command: python /path/to/ai-memory-layer/src/mcp_server.py

Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "ai-memory-layer": {
      "command": "python",
      "args": ["/absolute/path/to/ai-memory-layer/src/mcp_server.py"]
    }
  }
}

---

📦 Python SDK

Integrate the memory layer directly into your Python workflows or CI/CD pipelines.

from sdk import MemoryClient

client = MemoryClient(base_url="http://localhost:8000", api_key="your-secret-key")

# Ingest a repository history
client.ingest(repo_path="./my-project", project_id="my-app", max_commits=100)

# Recall architectural decisions
memories = client.recall("How do we handle auth?", project_id="my-app")
for m in memories:
    print(f"[{m['module']}] {m['content']}")

---

✨ Core Features

• Smart Deduplication: SHA256 content hashing prevents redundant memories.
• Conflict Detection: AI automatically flags if a new decision contradicts a previous one.
• Advanced MCP Tools: recall_memory, store_memory, list_recent_memories, flag_contradiction.
• Memory Dashboard: Built-in React UI at /dashboard with coverage heatmaps.

---

🗺️ Roadmap

• GitHub Actions Integration: Auto-ingest memories on every PR merge.
• Multi-User RBAC: Granular permissions for team-wide memory layers.
• Graph-Based Recall: Linking related decisions across different modules.
• Slack/Discord Bot: Capture decisions directly from team chats.

---

❓ Troubleshooting

• Database Connection Error: Ensure Docker is running and the port 5433 is not occupied.
• Embedding Failures: If using local, ensure you have enough RAM for the sentence-transformers model.
• MCP Not Loading: Ensure you use the absolute path to mcp_server.py in your agent configuration.

---

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md.

📄 License

MIT License - see LICENSE.