Skip to content

Synthesis: Knowledge Infrastructure for AI Agents

Synthesis is a local-first knowledge infrastructure platform that makes a codebase — and everything around it — navigable by AI agents. It indexes code, documents, PDFs, and even videos into multi-layer knowledge graphs, tracks how that knowledge changes over time, and exposes all of it through a CLI, an MCP server, and an LSP. Everything runs on your machine; nothing is sent to a cloud service.

It exists because of a problem I did not anticipate.

The problem Synthesis solves

In January 2026 I built lib-pcb — 197,831 lines of Java in eleven days, with an AI agent writing most of it. The methodology worked. And then I couldn't navigate any of it. Nearly 18,000 files. 691 new files per day. I was spending more time finding code than understanding it.

That is not a lib-pcb problem. It is the comprehension bottleneck: AI made creating easy but understanding harder. Output velocity jumped 10–50×; comprehension did not. Every jump in creation speed eventually produces a navigation crisis — monks didn't need a library catalog until the printing press arrived. I had given myself a personal printing press for code, and I was drowning in the output.

Synthesis started as search and became the missing layer. The full story: The Tool I Didn't Plan to Build: Synthesis, Ten Weeks Later.

At a glance

Local-first · indexes at 200–300 files/second · sub-second search · multi-layer knowledge graphs · 60+ CLI commands · 11 MCP tools · 4,300+ tests · Notion as a first-class workspace source · Java 21 · github.com/exoreaction/Synthesis

Agents need three kinds of memory

Most AI agents have exactly one kind of memory — the context window — and forget everything the moment the session ends. Synthesis gives them three:

  1. Working memory — the context window (the model already has this).
  2. Episodic memory — what happened in past sessions, indexed and searchable.
  3. Semantic memory — the workspace knowledge graph: what the code means and how it connects.

Read Working Memory, Episodic Memory, Semantic Memory — Your Agent Has One, then Your AI Has One Layer. It Needs Four. for the retrieval architecture.

And the uncomfortable follow-up: building the layers is the easy part. Memory that is not maintained becomes memory that lies — Agent Memory Rots. Here's How We Stopped It.

What it found by looking at itself

The most convincing evidence for a knowledge tool is what it surfaces when you point it at real, messy history.

Post Finding
What Synthesis Found in 31 Seconds An XXE vulnerability in a production Java SSO system — in 31 seconds.
The Synthesis Excavation Text coverage 99.6%, real asset coverage 15.2%. 4,852 binary files recovered from 3.5 years of lost history in one day.
Zero Links: An Engineering Session 777 directories, zero edges → 11,777 edges, 23 new tests, 4 bugs fixed, one day later.
We Gave the AI Better Documentation. It Got Slower. CLI docs increased tool calls 11%; MCP cut them 35%. How agents integrate matters more than how much you document.
The Mirror Test Using an AI tool to measure whether an AI tool can be trusted — the dogfooding loop.

Reading guide

Synthesis is under active development. The blog has the latest; this page has the map.