Score memories in first 60% of conversation by tokens

Use cumulative token position instead of entry index for the scoring
cutoff. This reflects actual context usage — a few large entries
near the end won't skew the boundary.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

src/subconscious/learn.rs

@@ -337,16 +337,25 @@ where
     let http = http_client();
     let mut scored = 0;
 
-    let total_entries = context.conversation.entries().len();
+    let total_tokens = context.conversation.tokens();
-    let first_quarter = total_entries / 4;
+    let token_cutoff = total_tokens * 60 / 100;
+
+    // Precompute cumulative token position for each entry
+    let entries = context.conversation.entries();
+    let mut cumulative: Vec<usize> = Vec::with_capacity(entries.len());
+    let mut running = 0;
+    for e in entries {
+        running += e.tokens;
+        cumulative.push(running);
+    }
 
     for (pos, key, _) in &candidates {
-        let (end, full_window) = nth_response_end(context.conversation.entries(), *pos, response_window);
+        // Only score memories in the first 70% of the conversation by tokens —
-        // Skip memories without a full window, unless they're in the
+        // recent memories don't have enough responses to evaluate yet.
-        // first quarter of the conversation (always score those).
+        if cumulative.get(*pos).copied().unwrap_or(total_tokens) > token_cutoff {
-        if !full_window && *pos >= first_quarter {
             continue;
         }
+        let (end, _) = nth_response_end(context.conversation.entries(), *pos, response_window);
         let range = *pos..end;
         if !context.conversation.entries()[range.clone()].iter().any(|ce| ce.entry.message().role == Role::Assistant) {
             continue;