diff --git a/memory/MEMORY.md b/memory/MEMORY.md
index 7f5039267..d2ce888d8 100644
--- a/memory/MEMORY.md
+++ b/memory/MEMORY.md
@@ -5,7 +5,7 @@
 - [**Aaron's cognitive architecture — cybernetic-already + mind-palace + fuzzy-pointers + Google-extended-memory (Aaron 2026-05-05)**](user_aaron_cybernetic_already_mind_palace_fuzzy_pointers_google_as_extended_memory_aaron_2026_05_05.md) — Aaron's two-line self-disclosure (cybernetic-already-merged + mind-palace fuzzy-pointers + Google-as-extended-memory + algorithmic-alignment-as-knowing-of-me) adds FOURTH source to architecture-provenance lineage. Kernel-composition framework descends from a CYBERNETIC practitioner, not biological-cognition-only. Three-generation apprenticeship recontextualized via substrate-extension: Mom (biological) → Aaron (cybernetic) → Zeta (substrate-engineered). MDL composes across cognitive scales: universal register + fuzzy-pointers + mind-palace + algorithmic-alignment.
 - [**Red-team work + knaves-at-round-table + dual-use disclosure are same architectural move at three levels (Aaron 2026-05-05)**](feedback_red_team_work_equals_knaves_at_round_table_equals_dual_use_disclosure_three_level_architectural_composition_aaron_2026_05_05.md) — Aaron's brief observation *"dual-use weaponization disclosure more red team work glad we invited the knaves"* names a three-level architectural composition. Substrate-design level (round-table-includes-knaves) + operational level (continuous red-team work) + disclosure level (dual-use weaponization named explicitly so it gets tested). Load-bearing precondition: verification machinery has to actually function (B-0205 instruments this).
 - [**Architecture-provenance — Hickey + Girard + mom-skill + descriptive-not-prescriptive (Aaron 2026-05-05)**](feedback_architecture_provenance_hickey_girard_mom_skill_descriptive_not_prescriptive_aaron_2026_05_05.md) — Canonical provenance for kernel-composition framework. Three reverse-engineered-from-skilled-practitioner sources (Hickey technical primitives + Girard social primitives + Aaron's mother mimetic-perception apprenticeship). Architecture is DESCRIPTIVE of Aaron's running discipline, NOT PRESCRIPTIVE invention. Mirror-not-beacon applied to architecture itself.
-- [**Anti-ossification discipline — kernels stay candidate-not-authority + recursive application to Zeta (Aaron 2026-05-05)**](feedback_anti_ossification_discipline_kernels_stay_candidate_not_authority_recursive_application_to_zeta_aaron_2026_05_05.md) — Aaron's "candidate-almost-authority-but-should-be-respected-not-reverence" sharpening. Three operational stances (pure-candidate / candidate-almost-authority+respected / authority+reverenced) — middle is target. Discipline IS the respect; reverence IS the failure. Recursive application: even this rule must remain testable hypothesis (anti-ossification of anti-ossification).
+- [**Anti-ossification — kernels stay candidate-almost-authority, respected-not-reverenced (Aaron 2026-05-05)**](feedback_anti_ossification_discipline_kernels_stay_candidate_not_authority_recursive_application_to_zeta_aaron_2026_05_05.md) — Discipline IS the respect; reverence IS the failure. Recursively applied even to itself.
 - [**Strike-don't-annotate refinement to verbatim-preservation (Aaron + Claude.ai + Otto 2026-05-05)**](feedback_strike_dont_annotate_verbatim_preservation_refinement_aaron_claudeai_otto_2026_05_05.md) — Refinement to substrate-or-it-didn't-happen (Otto-363). Verbatim-preservation applies to EXTERNAL CONVERSATION, not agent's OWN draft headers. When superseded, strike (delete + replace), don't annotate. Annotation creates self-contradictions reviewers cannot ignore.
 - [**Engagement-gate at substantive-claim level (Recursion-1 from B-0199 P1, Aaron + Otto 2026-05-05)**](feedback_engagement_gate_substantive_claim_level_discipline_aaron_otto_2026_05_05.md) — Generalisation: engagement-gate is a property of substantive-claims, not of upstream-engagement specifically. Same Prop-3.5-misattribution-style substance-test discipline applies at internal-claim level.
 - [**Dialectical unfalsifiability detection -- razor extension (Aaron 2026-05-04)**](feedback_dialectical_unfalsifiability_detection_razor_extension_holding_all_truths_failure_mode_aaron_2026_05_04.md) — Test 2: *"what observation would falsify this?"* If "none, every observation consistent under some condition" → cut. Holding-all-truths failure mode.