claimgraph: Structured Memory for Coding Agents

A bi-temporal, epistemically typed knowledge graph for coding-agent memory: the problem and the field, the design, the working system chapter by chapter, the benchmark, and the reference.

Preface

claimgraph is a memory system for AI coding agents. It stores what an agent and its human learn about a codebase (decisions, preferences, dependencies, failure modes, history) as a knowledge graph rather than as a pile of markdown files. Every fact in the graph knows when it was true, when it was recorded, how confident anyone should be in it, what kind of claim it is, and where it came from. Nothing is ever deleted: when the world changes, the old fact’s validity interval closes and the new one opens, so the graph can answer both “what do we believe now” and “what did we believe in March, and why did it change.”

The system is a Babashka CLI backed by Datalevin, wrapped in an agent skill, an MCP server, and a set of hooks that let it run with zero effort on the user’s part. This book covers the project’s background and rationale, its design, the benchmarks, and how to use it.

How to read this book

The book has three parts.

Part I — Foundations is prose. It explains the problem agent memory is trying to solve, what the research literature settled in 2025 and 2026, and the mental model behind claimgraph’s design. If you read nothing else, read the mental model chapter. It’s the foundation for understanding everything else.

Part II — The System in Practice is a hands-on tour of the working system, one behavior per chapter. These chapters use the in-memory store backend, which shares every line of decision logic with the Datalevin backend through a storage protocol. The CLI equivalents appear alongside as shell blocks.

Part III — Operations and Reference is operational: advanced usage, a hands-on guide to extending the code tier to a new language (and merging that work upstream), the benchmark and its results, a comparison with the other memory systems in the field as of July 2026, a CLI reference, and the bibliography.

Building the book

The rendered book is generated from book/ in the repository:

bb book            # render to book/rendered/_book/index.html
bb book:preview    # render, then serve with quarto preview

The build needs a JVM (the book chapters evaluate on real Clojure, not on Babashka) and the Quarto CLI. The claimgraph tool itself needs neither; it runs on two native binaries.

Status

Everything described here is implemented and tested. There is also a new benchmark included in the product, with 33 questions about a made-up project. All pass and gate regressions in CI. The end-task A/B and its numbers appear in the benchmark chapter, including the arms where claimgraph loses and the one where the best available answer is “the graph does not know.”