Koragraph Research
Our research.
The experiments behind our claims, released with the corpus and the oracles so anyone can re-run them against us.
Papers
4 papers54Humans
Self-hosted memory for coding agents
A 1.7B-parameter local model that is the only writer of a coding agent’s memory raised cross-session rule compliance from 63.0% with Claude Code’s built-in memory to 94.5%, and delivered rules were followed 95.0% of the time against 28.4% for rules that never arrived.
94.5%
later-session rule compliance, against 63.0% with built-in memory
9/10
constraints recorded, against 4/10 by the agent itself
95.0%
of delivered rules followed, against 28.4% of undelivered
Read the paper →
The Authorship Floor
Why project knowledge is lost at both ends, and what anchoring it to code recovers
Project documentation fails at both ends at once: it never captures what is too small to write down, and it rots what is large enough to write, and both failures share one cause. We describe the memory layer we built on the argument that the two costs have to fall together, and we do not claim it improves outcomes.
2
documentation failures, below the floor and above it, from one cause
6
verdicts per anchor at index time; one withdraws, one changes nothing
unknown
the verdict returned under nine conditions rather than guessing
Read the paper →
94% of Frontier Retrieval Quality from a 494M-Parameter Model
Method summarisation for code knowledge graphs in 2.43 GPU-hours, and the point where more training data stops paying
A 494M-parameter model fine-tuned for 2.43 T4-hours, about $0.85 of commodity GPU time, writes the method descriptions a code knowledge graph is searched on at 93.8% of the frontier teacher’s retrieval quality, and beats a hosted model Koragraph itself was paying for.
93.8%
of frontier teacher retrieval quality, hit@8
$0.85
commodity GPU time to train the shipping model
91%
of the total gain from the first 2,500 examples
Read the paper →
Can Graph-Based Context Engineering Substitute for Model Scale?
Task Resolution and Cost for Cheap and Frontier Models on Repository-Level Code Tasks
A cheap model given graph-derived context solved 9 of 11 repository-level tasks at $0.027 per solved task. A frontier agentic system searching the same repositories with grep solved 9 as well, at $2.82.
2 → 9
cheap tier, tasks solved, grep → graph
9 → 9
frontier tier, tasks solved, grep → graph
105.9×
cost per solved task, frontier+grep vs cheap+graph
Read the paper →
What is next
A larger corpus, larger repositories, more languages. The task suite is built to grow, and independent replication is the reason it is public.