Qontext: A Quipu-Inspired Conversational Memory for Local Language Models

Abstract

A conversational memory for language models that stores short, self-contained third-person facts ("knots") and returns only those a given prompt needs, inspired by the Inca quipu, AI use fully disclosed in the paper's Author contributions section. On 800-turn conversations with human-written dialogue as filler, a 12B model answers 9.7/10 planted-fact questions from a 116-token pack against 9.3/10 from a 13,853-token transcript: tied accuracy at a 119-fold reduction in prompt tokens. On smaller models the reduction also improves accuracy, but the effect vanishes at 12B, so it is stated conditionally. VERSION 2 RETRACTS A SUBSTANTIAL PART OF VERSION 1. Version 1 reported a decomposition of retrieval failure on conversational-turn queries (79.8% of misses unreachable) and six retrieval mechanisms measured against it. All of it was computed from a benchmark whose ground truth was inferred from each conversation's own continuation. That benchmark fails a shuffle control: pairing a turn with the wrong reply marks 92% as many facts "needed" as the right one, and a reply from an entirely unrelated conversation marks 97%. The same control separates correct from incorrect by 42.9x on the paper's hand-keyed suites, so the instrument works and the benchmark does not. Every turn-shaped and roleplay result is withdrawn. The cost and accuracy results use hand-written answer keys, share no instrumentation with the retracted section, and are unchanged. Section 6 now reports the failure, the control that detects it, and an explicit list of what is withdrawn — on the argument that a proxy ground truth must be shown to fail when the answer is wrong, and that this check is almost never published. Code, benchmarks and all measurements are released, including five occasions on which the authors' own instrumentation produced false findings.