WriterLore

A personal, human-curated semantic memory layer for AI-assisted documentation work. Built as a custom MCP server backed by MongoDB Atlas.

What it does

Every Claude Code session starts from zero. Context about editorial decisions, structural patterns, and cross-repo choices lives in markdown files that don't scale and can't reason about relevance. WriterLore lets you save a decision once and surface it automatically in future sessions — in any repo.

It is not an automatic capture system. You decide what is worth keeping.

How it works

WriterLore exposes two MCP tools to Claude Code:

• writerlore_save — explicitly save an editorial decision, pattern, or workflow quirk to MongoDB. The current repo and Jira ticket are auto-detected from git remote and the branch name. If no ticket is found in the branch name, Claude looks it up in Jira before saving.
• writerlore_recall — semantic search across saved memories, optionally filtered by repo.

At session start, a hook queries MongoDB and injects the top 3 relevant memories into the session context automatically.

Stack

Component	Choice
Language	Python 3.9+
Database	MongoDB Atlas — writerLore.memories
Embeddings	all-MiniLM-L6-v2 via sentence-transformers (384 dims, local)
MCP server	FastMCP (mcp>=1.0)
MCP config	~/.mcp.json

Prerequisites

• Python 3.9+
• A MongoDB Atlas cluster (free tier works)
• Claude Code CLI

Setup

git clone https://github.com/xargom/writerlore
cd writerlore
bash setup.sh

setup.sh will:

1. Create a .venv and install dependencies
2. Register the MCP server in ~/.mcp.json (prompts for your Atlas connection string)
3. Register the session hook in ~/.claude/settings.json
4. Print the Atlas Vector Search index definition to create manually

Restart Claude Code after setup completes.

Atlas Vector Search index

Create a vector index named vector_index on the memories collection in the writerLore database:

{
  "fields": [
    { "type": "vector", "path": "embedding", "numDimensions": 384, "similarity": "cosine" },
    { "type": "filter", "path": "primary_repo" },
    { "type": "filter", "path": "also_relevant_for" },
    { "type": "filter", "path": "components" }
  ]
}

Memory schema

{
  "content": str,              # The editorial decision or pattern
  "primary_repo": str,         # Auto-detected from git remote
  "also_relevant_for": [str],  # Other repos this applies to
  "ticket": str | None,        # Auto-detected from branch name, or Jira lookup
  "components": [str],           # e.g. ["Atlas Search", "Vector Search"]
  "saved_at": datetime,        # UTC
  "embedding": [float],        # 384-dim local sentence-transformers vector
}

Usage

Save a memory — say this to Claude during a session:

"Save to WriterLore: In Terraform module repos, always use absolute GitHub URLs for cross-file anchors because the template renders at two different directory depths."

Recall on demand:

"What do I know about PR workflow for Terraform repos?"

Automatic recall happens at every session start — no action needed.

Known limitations

• session-recall.py runs as a standalone hook and cannot reach MCP — ticket detection there is branch-name only.
• Memories are personal and not team-shared (v1 scope).
• The local model (all-MiniLM-L6-v2, ~90 MB) is downloaded on first use.