<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0">
<channel>
  <title>HeurChain Research</title>
  <link>https://heurchain.com/research</link>
  <description>Paper-linked research notes and benchmark results from the HeurChain team.</description>
  <language>en</language>
  <item>
    <title>The memory retrieval advantage: Recursive Improvement</title>
    <link>https://heurchain.com/research/s2-memory-retrieval-advantage</link>
    <guid isPermaLink="true">https://heurchain.com/research/s2-memory-retrieval-advantage</guid>
    <pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate>
    <description>With no memory, three readers answer 0 of 543 LoCoMo DEV questions. Retrieval version S2 adds 33 to 38 correct answers per reader and passed a sealed holdout check (net +50, p &lt; 0.0001). Recursive improvement then found p04a, which beats S2 on a second holdout check (net +22, p = 0.004) and Mem0 for every reader.</description>
    <category>benchmarking</category>
    <category>locomo</category>
    <category>memory-systems</category>
    <category>retrieval</category>
  </item>
  <item>
    <title>Six memory systems compared on LoCoMo DEV</title>
    <link>https://heurchain.com/research/six-systems-locomo-dev</link>
    <guid isPermaLink="true">https://heurchain.com/research/six-systems-locomo-dev</guid>
    <pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate>
    <description>Two independent runs per system on four LoCoMo DEV conversations (543 scored questions; adversarial category 5 excluded): HeurChain 70.6%, HeurChain with date writing 72.5%, Mem0 71.3%, Hindsight 48.8%, Zep and OpenViking 37.7%. The top-three gap sits inside run-to-run noise. DEV split only.</description>
    <category>benchmarking</category>
    <category>locomo</category>
    <category>memory-systems</category>
  </item>
  <item>
    <title>Why temporal answers fail on evidence already retrieved</title>
    <link>https://heurchain.com/research/temporal-multihop-miss-taxonomy</link>
    <guid isPermaLink="true">https://heurchain.com/research/temporal-multihop-miss-taxonomy</guid>
    <pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate>
    <description>Of HeurChain's 49 temporal misses on the LoCoMo DEV split, 36 had every labelled evidence turn already packed (an answer-side failure over unresolved relative dates), while 45 of 55 multi-hop misses lacked evidence (a retrieval-coverage failure). HeurChain's 17-question single-hop lead over Mem0 is the lead to protect.</description>
    <category>benchmarking</category>
    <category>locomo</category>
    <category>temporal-reasoning</category>
    <category>multi-hop-reasoning</category>
    <category>retrieval</category>
  </item>
</channel>
</rss>
