Let the assistant propose what to remember, and the operator decide what is kept

Deployment/master.conf.template

@@ -1762,6 +1762,23 @@
 # truncated away — the exact inversion of which one matters.
     AI_MEMORY_LEARNED_MAX_CHARS=1200
 
+# ━━━ AI Learned Memory ━━━
+# Lets the assistant propose facts worth keeping. It never writes memory directly: a candidate is
+# filed in data/ai/mem_proposals/ and the operator accepts or dismisses it, exactly as findings
+# work. Accepted text lands in the learned slot, which the prompt ranks BELOW retrieval.
+#
+# The reason for the indirection: memory rides on every future prompt. A model allowed to write
+# its own — and told to prefer it over the passages — would restate a wrong conclusion forever,
+# reading its own claim back as evidence. A bad proposal must cost one dismissal, not that.
+#
+# Costs nothing while off: the instruction that asks for a candidate is only added to the prompt
+# when this is true, so a disabled feature is genuinely absent rather than merely ignored.
+    AI_MEMORY_LEARN_ENABLED=false
+
+# Writes accepted candidates without asking. Cannot outrank its parent — with proposing off this
+# does nothing. Leave it false until the proposals have proven good for a while.
+    AI_MEMORY_LEARN_AUTO_ACCEPT=false
+
 # ━━━ AI Stored Conversations ━━━
 # How many past conversations the AI tab and the Monitor tab's AI row keep. One JSON file per
 # conversation under DATA_DIR/ai_chats, saved automatically when a turn completes; the oldest