Memory Anchor

The Problem

# Monday — Session 1
You: "Let's use Qdrant for vector search, SQLite for metadata."
AI:  "Great choice. I'll structure the project around that."

# Tuesday — Session 2
You: "Continue where we left off."
AI:  "I don't have any context about previous sessions."
You: 😐

# Tuesday — Session 2, with Memory Anchor
You: "Continue where we left off."
AI:  [searches memory] "Got it — we decided on Qdrant + SQLite yesterday.
      I also see 3 open bugs from last session. Want to start there?"

Memory Anchor is an MCP server that gives AI coding assistants persistent, searchable memory. No cloud, no API keys, no Docker — just pip install and go.

5-Minute Setup

# Install from GitHub
pip install git+https://github.com/bao243092078-crypto/memory-anchor.git

# Configure Claude Code MCP integration
memory-anchor setup --project my-project

# Restart Claude Code — done!

That's it. SQLite is embedded. Everything stays on your machine.

What Makes This Different

MCP Tools

3-Layer Memory Model

L0  Identity     — Who you are. Protected by 3-approval mechanism.
L2  Events       — What happened. Timestamped, optional TTL.
L3  Facts        — What's true. Verified long-term knowledge.

How It Works

Claude Code <-> MCP (stdio) <-> Memory Kernel <-> SQLite + Full-Text Search
                                      |
                               Cold Storage (JSONL backup)

• Smart search: Finds what you mean, not just exact words (vector + full-text + keyword overlap)
• Auto-categorization: Memories are tagged and layered automatically
• Write safety: Validates confidence before storing — nothing gets lost silently
• Identity protection: Core identity requires 3 separate approvals to modify
• Backup: Append-only JSONL log means memories survive even if the database breaks
• 100% local: Zero data leaves your machine

Use Cases

Long-running projects — After 50+ sessions, your AI still remembers why you chose Qdrant over Pinecone, which auth approach you abandoned, and what the deployment strategy is.

Bug fix tracking — "How did we fix that null pointer last week?" Your AI searches memory and finds the exact commit, root cause, and fix.

Cross-project patterns — Decisions from project A are searchable from project B. Each project has its own namespace, but you can query across them.

Configuration

# Each project gets its own memory namespace
memory-anchor serve --project project-a
memory-anchor serve --project project-b

Manual MCP config (if you prefer not to use memory-anchor setup):

{
  "mcpServers": {
    "memory-anchor": {
      "command": "memory-anchor",
      "args": ["serve", "--project", "my-project"]
    }
  }
}

CLI

memory-anchor setup  --project NAME   # Configure Claude Code MCP
memory-anchor serve  --project NAME   # Start MCP stdio server
memory-anchor doctor --project NAME   # Health check
memory-anchor init   --project NAME   # Initialize project config

Development

git clone https://github.com/bao243092078-crypto/memory-anchor.git
cd memory-anchor
uv sync --all-extras

uv run pytest                         # Run all tests
uv run ruff check backend --fix       # Lint
uv run mypy backend                   # Type check

License

MIT License