What it is
AKG (Agentic Knowledge Graph) is a portable binary graph format built specifically for agent memory. Think of it as SQLite for agent memory: a compact, file-based way to store and query knowledge graphs with no server, no setup, and no dependencies.
It ships with a conformance spec and reference fixtures to keep implementations consistent across languages, and official SDKs in Go and TypeScript for embedding AKG-backed memory into agentic development workflows and user-facing AI products.
Why I built it
Most agent memory solutions are either too heavy (vector databases, hosted services) or too fragile (stuffing everything into the context window and hoping for the best). I wanted something local-first, deterministic, and portable, something you can ship alongside your application the same way SQLite ships with millions of apps.
I found a graph storage pattern I liked, noticed it had real limitations at scale, and decided the only way to understand binary file storage well enough to fix them was to go build it from scratch. A week of diving into how SQLite and RocksDB handle their on-disk formats later, AKG had its first working spec.
How it works
AKG stores nodes and edges in a compact binary layout optimized for agent use cases: fast lookups by node ID, efficient traversal of relationship graphs, and a small file footprint even for large knowledge bases. The format is append-friendly, making it practical for agents that update memory incrementally rather than rewriting it wholesale.
The Go and TypeScript SDKs expose a simple read/write API that maps naturally onto the kinds of memory operations agents actually need: storing facts, retrieving context, updating beliefs, and querying relationships.
Status
The core format spec is stable and the Go and TypeScript SDKs are under active development.