perf(memory): defer to_lowercase in extract_facts until triggers fire

CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## Unreleased
 
+### Memory
+
+- **Auto-capture**: defer the allocating `to_lowercase` in `extract_facts` until
+  an identity, preference, or relationship trigger is present — common no-match
+  utterances (weather, questions, chit-chat) skip the allocation entirely.
+
 ## 1.0.0-rc.3 - 2026-07-03
 
 Third release candidate. This RC lays the **M2 foundations** — a portable

crates/genie-core/src/memory/extract.rs

@@ -24,6 +24,20 @@ pub struct ExtractedFact {
 pub fn extract_facts(text: &str) -> Vec<ExtractedFact> {
     let mut facts = Vec::new();
     let trimmed = text.trim();
+
+    if !needs_extract_facts_lower(text) {
+        if trimmed.len() >= 8
+            && trimmed[..8].eq_ignore_ascii_case("remember")
+            && let Some(content) = extract_remember(trimmed)
+        {
+            facts.push(ExtractedFact {
+                category: "fact".into(),
+                content,
+            });
+        }
+        return facts;
+    }
+
     let lower = text.to_lowercase();
 
     if needs_identity_scan(&lower) {
@@ -193,6 +207,26 @@ pub fn extract_and_store(memory: &Memory, user_text: &str) -> usize {
 
 // --- Pattern helpers ---
 
+/// ASCII case-insensitive substring check for pre-lowercase trigger gates.
+fn contains_ascii_ci(haystack: &str, needle: &str) -> bool {
+    if needle.is_empty() || haystack.len() < needle.len() {
+        return false;
+    }
+    haystack.as_bytes().windows(needle.len()).any(|window| {
+        window
+            .iter()
+            .zip(needle.bytes())
+            .all(|(left, right)| left.eq_ignore_ascii_case(&right))
+    })
+}
+
+/// True when any Tier-1 extractor needs the allocating lowercase view.
+fn needs_extract_facts_lower(text: &str) -> bool {
+    needs_identity_scan_ci(text)
+        || needs_preference_scan_ci(text)
+        || needs_relationship_scan_ci(text)
+}
+
 /// Phrases that mark where a captured value ends. A fact like "my name is X"
 /// must capture only X, not the conjunction or subordinate clause that follows
 /// it ("...and I love coding", "...but I hate meetings", "...who lives nearby").
@@ -282,32 +316,44 @@ fn extract_age(text: &str) -> Option<u32> {
 }
 
 fn needs_identity_scan(lower: &str) -> bool {
-    lower.contains("my name")
-        || lower.contains("call me")
-        || lower.contains("called")
-        || lower.contains("i work")
-        || lower.contains("working at")
-        || lower.contains("work as")
-        || lower.contains("i live")
-        || lower.contains("i'm from")
-        || lower.contains("i am from")
-        || lower.contains("based in")
-        || lower.contains("i'm a")
-        || lower.contains("i am a")
-        || lower.contains("i'm ")
-        || lower.contains("i am ")
+    needs_identity_scan_ci(lower)
+}
+
+fn needs_identity_scan_ci(text: &str) -> bool {
+    contains_ascii_ci(text, "my name")
+        || contains_ascii_ci(text, "call me")
+        || contains_ascii_ci(text, "called")
+        || contains_ascii_ci(text, "i work")
+        || contains_ascii_ci(text, "working at")
+        || contains_ascii_ci(text, "work as")
+        || contains_ascii_ci(text, "i live")
+        || contains_ascii_ci(text, "i'm from")
+        || contains_ascii_ci(text, "i am from")
+        || contains_ascii_ci(text, "based in")
+        || contains_ascii_ci(text, "i'm a")
+        || contains_ascii_ci(text, "i am a")
+        || contains_ascii_ci(text, "i'm ")
+        || contains_ascii_ci(text, "i am ")
 }
 
 fn needs_preference_scan(lower: &str) -> bool {
-    lower.contains("i like")
-        || lower.contains("i love")
-        || lower.contains("i enjoy")
-        || lower.contains("i prefer")
-        || lower.contains("i hate")
-        || lower.contains("i dislike")
-        || lower.contains("i don't like")
-        || lower.contains("i can't stand")
-        || lower.contains("favo")
+    needs_preference_scan_ci(lower)
+}
+
+fn needs_preference_scan_ci(text: &str) -> bool {
+    contains_ascii_ci(text, "i like")
+        || contains_ascii_ci(text, "i love")
+        || contains_ascii_ci(text, "i enjoy")
+        || contains_ascii_ci(text, "i prefer")
+        || contains_ascii_ci(text, "i hate")
+        || contains_ascii_ci(text, "i dislike")
+        || contains_ascii_ci(text, "i don't like")
+        || contains_ascii_ci(text, "i can't stand")
+        || contains_ascii_ci(text, "favo")
+}
+
+fn needs_relationship_scan_ci(text: &str) -> bool {
+    contains_ascii_ci(text, "my ")
 }
 
 fn extract_favorite(text: &str) -> Option<String> {

crates/genie-core/tests/extract_bench.rs

@@ -21,8 +21,8 @@ fn run(label: &str, input: &str, iters: u32) {
 #[test]
 #[ignore = "benchmark; run with --release --ignored --nocapture"]
 fn bench_extract_facts() {
-    // Common case: no relationship, identity, or preference phrase — skips all
-    // prefix scans after the single to_lowercase (see #495 early-outs).
+    // Common case: no relationship, identity, or preference phrase — skips the
+    // allocating to_lowercase entirely (see #495 early-outs + deferred lower).
     run(
         "no-match",
         "the weather today is nice and i went for a walk",

crates/genie-core/tests/extract_test.rs

@@ -10,7 +10,8 @@ fn facts_key(facts: &[ExtractedFact]) -> String {
 }
 
 /// Fixed corpus captured from `main` @ 72b0b28 — guards byte-identical output after
-/// identity/preference early-outs and favorite/remember allocation trims (#495).
+/// identity/preference early-outs, favorite/remember allocation trims (#495), and
+/// deferred `to_lowercase` on no-match utterances.
 #[test]
 fn extract_facts_corpus_regression() {
     const CORPUS: &[(&str, &str)] = &[
@@ -80,6 +81,8 @@ fn extract_facts_corpus_regression() {
         ("my plan today is to relax and read a good book", ""),
         ("REMEMBER my pin is 1234", "fact:my pin is 1234"),
         ("Remember that", ""),
+        ("WHAT TIME IS IT?", ""),
+        ("The Weather Today Is Nice", ""),
     ];
 
     for (input, expected) in CORPUS {