self-review 2026-05-18: raise SEM_FLOOR to 0.70, add ACT-R frequency resistance to decay

Semantic seed floor raised 0.65 → 0.70: field literature (SuperLocalMemory
arXiv:2604.04514) validates 0.70-0.75 as the noise floor for sentence-transformer
embeddings on non-trivial corpora. The 0.65 threshold was admitting false
positives that diluted BFS traversal quality on the 13K-node post-ingestor graph.
Top-30 cap retained — sparse graph (1.26 edges/node) needs more semantic entry
points than a dense graph would.

ACT-R Base-Level Learning insight applied to engram_temporal_decay: current
purely time-based formula treated a node activated 100 times identically to a
node never activated (same decay rate). ACT-R BLL (Anderson 1993) shows
frequently-accessed memories resist temporal decay. Fix: freq_resist multiplier
= 1.0 + 0.1 * log(1 + activation_count). count=0→1.0×, count=9→1.23×,
count=99→1.46×. Existing activation_dampening continues to reduce per-query
novelty for well-known nodes — the two mechanisms are complementary: resist
decay (durable in graph), dampen per-query (don't dominate any single turn).

lang/releases/v1.0.0-20260501/el_runtime.c

@@ -6552,8 +6552,21 @@ el_val_t engram_edge_count(void) {
 }
 
 /* Compute temporal decay factor for a node given current time.
- * effective contribution = salience * exp(-lambda * age_hours / T_half)
- * Clamped to [0.05, 1.0] so very old nodes retain a meaningful floor. */
+ * effective contribution = salience * exp(-lambda * age_hours / T_half) * freq_resist
+ * Clamped to [0.05, 1.0] so very old nodes retain a meaningful floor.
+ *
+ * ACT-R-inspired frequency resistance (added 2026-05-18 self-review):
+ *   freq_resist = 1.0 + 0.1 * log(1 + activation_count)
+ *   count=0 → 1.00×, count=9 → 1.23×, count=99 → 1.46×, count=999 → 1.69×
+ *
+ * Rationale: ACT-R Base-Level Learning (Anderson 1993, d=0.5 power law) shows
+ * that frequently-accessed memories resist temporal decay — the brain doesn't
+ * let well-rehearsed knowledge fade at the same rate as never-accessed nodes.
+ * Our activation_dampening already reduces per-query novelty for high-count
+ * nodes (correct — prevents runaway dominance in any single turn). But without
+ * freq_resist, that same node was decaying as fast as a node never activated,
+ * which is cognitively wrong. This multiplier corrects it: canonical nodes and
+ * heavily-used lessons resist decay, session-ephemeral nodes do not. */
 static double engram_temporal_decay(const EngramNode* n, int64_t now_ms) {
     int64_t age_ms = now_ms - n->last_activated;
     if (age_ms <= 0) return 1.0;
@@ -6561,6 +6574,10 @@ static double engram_temporal_decay(const EngramNode* n, int64_t now_ms) {
                                                     : ENGRAM_DECAY_LAMBDA;
     double age_hours = (double)age_ms / 3600000.0;
     double factor = exp(-lambda * age_hours / ENGRAM_T_HALF_HOURS);
+    /* Frequency resistance: frequently-activated nodes decay more slowly. */
+    double freq_resist = 1.0 + 0.1 * log(1.0 + (double)n->activation_count);
+    factor *= freq_resist;
+    if (factor > 1.0) factor = 1.0;
     if (factor < 0.05) factor = 0.05;
     return factor;
 }
@@ -6758,11 +6775,16 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
     }
     /* ── Semantic seed supplement ────────────────────────────────────────────
      * When the query has an embedding vector, scan un-seeded nodes for cosine
-     * similarity ≥ 0.65 and inject the top-30 as additional Pass-1 seeds.
+     * similarity ≥ 0.70 and inject the top-30 as additional Pass-1 seeds.
      * This ensures semantically-related nodes that share no query tokens still
      * enter the BFS fan-out.  Sem seeds get 85% of full act to stay slightly
      * weaker than exact lexical matches.  Top-K=30 cap prevents queue flooding
-     * on large graphs (11K+ nodes post-ingestor).
+     * on large graphs (13K+ nodes post-ingestor).
+     *
+     * Floor raised from 0.65 → 0.70 (2026-05-18 self-review): field literature
+     * (SuperLocalMemory, floop) validates 0.70-0.75 as the noise floor for
+     * nomic-embed-text; 0.65 was admitting false positives that diluted BFS
+     * traversal quality.
      *
      * Without this pass, embeddings collected at write-time via engram_embed_node
      * were only used in goal_bias (Pass 2) for nodes that already had lexical
@@ -6770,7 +6792,7 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
      * completely invisible to activation regardless of embedding quality. */
     if (qvec && qdim > 0) {
         typedef struct { int64_t idx; float sim; } SemEntry;
-        const float SEM_FLOOR = 0.65f;
+        const float SEM_FLOOR = 0.70f;
         SemEntry    top_sem[30];
         int64_t     top_sem_cnt = 0;
         float       top_sem_min = SEM_FLOOR;