From 8ce9371e56386ec9ccce003b9fda146a7db13d34 Mon Sep 17 00:00:00 2001
From: Ben <me@benjaminlee.co.uk>
Date: Mon, 3 Aug 2026 18:45:20 +0100
Subject: [PATCH] fix: exempt canonical corpus chunks from length normalization
 and decay
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Canonical corpus recall was near-impossible: after fusion, results pass
through applyLengthNormalization and applyDecayBoost/applyTimeDecay,
both calibrated for conversation memories.

- Corpus chunks are line-span document chunks (~3,900 chars against the
  500-char anchor -> factor ~0.40). Length is a property of the chunker,
  not entry quality, and chunk size is already bounded by the indexer.
- Corpus rows carry the source file's mtime as their timestamp
  (toMemoryEntry: timestamp: doc.mtimeMs), so weeks-old reference docs
  take the decay multiplier down to ~boostMin. Reference truth does not
  age like chat memory.

Measured: a corpus chunk that fused at 0.917 (top FTS hit, exact-keyword
floor) finished ~0.15 after both stages — below hardMinScore/minScore —
so memory_search and memory_recall returned 0 results for content that
is the store's #1 match on both retrieval sides. With the exemption the
same query returns it correctly.

Corpus ids are always 'corpus:'-prefixed (buildCorpusId), so the guard
is a cheap string check with no metadata parsing.
---
 src/retriever.ts | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/src/retriever.ts b/src/retriever.ts
index 2d69c81ca..755922a8c 100644
--- a/src/retriever.ts
+++ b/src/retriever.ts
@@ -1634,7 +1634,11 @@ export class MemoryRetriever {
 
     const reranked = results.map((result, index) => ({
       ...result,
-      score: clamp01(scored[index].score, result.score * 0.3),
+      // Corpus rows carry the source file's mtime as their timestamp;
+      // reference material must not decay like conversation memory.
+      score: (result.entry.id ?? "").startsWith("corpus:")
+        ? result.score
+        : clamp01(scored[index].score, result.score * 0.3),
     }));
 
     return reranked.sort((a, b) => b.score - a.score);
@@ -1654,6 +1658,11 @@ export class MemoryRetriever {
     if (!anchor || anchor <= 0) return results;
 
     const normalized = results.map((r) => {
+      // Canonical corpus chunks are line-span document chunks: their length is
+      // a property of the chunker, not of entry quality. Normalising them by
+      // length double-penalises reference material (chunk size is already
+      // bounded by the indexer).
+      if ((r.entry.id ?? "").startsWith("corpus:")) return r;
       const charLen = r.entry.text.length;
       const ratio = charLen / anchor;
       // No penalty for entries at or below anchor length.
@@ -1688,6 +1697,8 @@ export class MemoryRetriever {
 
     const now = Date.now();
     const decayed = results.map((r) => {
+      // Reference chunks keep file mtimes — do not age them.
+      if ((r.entry.id ?? "").startsWith("corpus:")) return r;
       const ts =
         r.entry.timestamp && r.entry.timestamp > 0 ? r.entry.timestamp : now;
       const ageDays = (now - ts) / 86_400_000;