ExtenGen: A Lightweight Shared Memory Layer for AI Coding Agents

Modern software teams increasingly rely on AI coding agents such as Claude Code, Cursor, and Codex to write, review, and debug source code. However, each agent session begins without knowledge of earlier sessions, so that decisions, architectural context, and institutional knowledge distributed across GitHub issues, pull requests, and chat threads are not carried forward. Developers consequently re-explain the same background repeatedly, which consumes time and tokens and produces inconsistent answers across team members. This report presents ExtenGen, a lightweight, self-hostable shared memory layer for AI coding agents. ExtenGen is implemented as a Model Context Protocol (MCP) server that exposes two tools, store_observation and recall_context, and it follows a five-stage pipeline of Capture, Embed, Store, Retrieve, and Inject. Captured text is converted into vector embeddings using the OpenAI text-embedding-3-small model and stored in PostgreSQL with the pgvector extension, while row-level security policies restrict every row to authorised organisations and users. At query time, a cosine-similarity search combined with an access-control filter returns the top-k most relevant authorised observations, which are merged into the agent's prompt. The lightweight version supports two ingestion sources, live MCP agent sessions and GitHub commit and pull requests, and is intended for small-team, self-hosted deployment. The project demonstrates Retrieval-Augmented Generation, vector similarity search, and multi-tenant access control in a single applied system and aims to reduce repeated context explanation and token usage.