Semantic Code Intelligence

The problem

Coding agents rediscover the same facts about a codebase on every session — where a symbol lives, who calls it, why a previous approach failed. That rediscovery is slow, expensive and lossy, and everything an agent learns dies with its context.

The solution

Source-code indexing with semantic search and structural code retrieval, served to coding agents over MCP, plus a persistent memory store so decisions and gotchas survive between sessions and between agents.

Technical implementation

An ongoing E-PATH tool, built because the problem kept getting in the way of the actual work: an AI coding agent that has to rediscover a codebase from scratch every session is doing expensive work twice.

Technical implementation

Two halves that are more useful together than apart.

Retrieval. Source code is indexed structurally and semantically, so an agent can ask both precise questions — where is this declared, who calls it, what breaks if I change it — and vague ones, like find the code that handles retry backoff when nobody remembers what it was named. Structural lookup answers the first kind exactly; embedding search answers the second kind well enough to be worth having.

Memory. A persistent store for the things an agent works out along the way: a decision and why it was taken, a failing command and the flag that fixed it, an expectation the codebase violated. Memories are attached to the symbols they concern, so they surface when that code is next touched rather than sitting in a document nobody opens.

Both are served over MCP, which means they are not tied to one agent or one vendor — any MCP-capable coding agent can use the same index and the same memories.

The honest status is ongoing. It is used daily on E-PATH’s own work, which is the point: the requirements come from real use rather than from a specification written up front.