PRemory

agent memory for org PR reviews

Every reviewer has a mental list of things they flag over and over that aren't bugs or lint errors — just "how we do things here." PRemory mines that history out of your team's merged pull requests and serves it back to coding agents as ranked, evidence-backed conventions, so an agent can check before writing code instead of you catching the same comment on the PR for the tenth time.

How It Works

1. Ingest: Merged PR review comments flow in automatically, either via a GitHub webhook on a repo, or via the VS Code extension's background sync (using your own GitHub login — no admin access needed) for whatever repo you have open.
2. Extract: The next time an agent asks a question, stale data is compiled on demand — raw comments are linked to the code they were left on, classified, and clustered into Conventions with supporting evidence.
3. Serve: An MCP server exposes that memory as tools an agent can call before or while writing code — ranked conventions for a path, similar historically-rejected patterns, and a predicted review of a diff you're about to propose.
4. Guard: A pre-commit hook and VS Code integration can flag staged changes that match a high-confidence historical rejection before you even open the PR.

See docs/ARCHITECTURE.md for the full data model and design rationale, and CLAUDE.md for a package-by-package tour of the codebase.

Quick Start

Prerequisites

• Node.js 22+ (see .nvmrc)
• Docker (for local PostgreSQL)
• A GitHub token with repo read access, for ingesting PR history
• Optional: ANTHROPIC_API_KEY / GEMINI_API_KEY for semantic extraction and LLM validation fallback — the deterministic analyzer works with no keys at all

Setup

npm install

# root env: GITHUB_TOKEN, DATA_DIR, optional model keys
cp .env.example .env

# api-server env: DATABASE_URL, API_HOST/PORT
cp packages/api-server/.env.example packages/api-server/.env

npm run db:setup       # start local Postgres + run migrations
npm run api-server     # GraphQL API on :8790

Get some memory into it

npm run ingest -- owner/repository   # bulk backfill from a repo's merged PR history
npm run extract                      # compile raw comments into conventions
npm run mcp-server                   # serve conventions over MCP

Or skip the manual steps and install the VS Code extension — it ingests and refreshes memory for whatever repo you have open, automatically, in the background.

Optional: continuous ingestion + pre-commit guard

npm run webhook-server    # merge-only GitHub webhook listener, for always-on ingestion on one repo
npm run hooks:install     # install the pre-commit review gate
npm run review-check      # manually validate the currently staged diff

Architecture

npm-workspaces monorepo. Every package talks to the others only through JSON files in data/ (or Postgres, in production) — never through direct cross-package imports — with @ht6/shared as the one exception for shared type contracts.

Package	Role
packages/ingestion	GitHub → raw review comments
packages/extraction	Raw comments → linked episodes → clustered conventions
packages/mcp-server	Retrieval, diff validation, and the 5 MCP tools agents call
packages/pipeline	Cross-stage freshness bookkeeping (ingest vs. extract staleness)
packages/webhook-server	Merge-only GitHub webhook listener
packages/vscode-extension	Editor integration: background sync, save-time warnings, pre-commit popup
packages/api-server	GraphQL API in front of Postgres, used by the extension and dashboard
dashboard	Read-only web view of conventions and review episodes

Development

npm run build       # tsc -b across all packages
npm run typecheck   # full rebuild, surfaces type errors
npm run test         # vitest across all packages

See docs/PIPELINE.md and docs/DATA_FORMAT.md for the stage-by-stage data flow and interchange file shapes.