diff --git a/src/aelfrice/feedback.py b/src/aelfrice/feedback.py
index 0d605e8a1..174a12343 100644
--- a/src/aelfrice/feedback.py
+++ b/src/aelfrice/feedback.py
@@ -85,6 +85,7 @@ def apply_feedback(
     source: str,
     now: str | None = None,
     propagate: bool = True,
+    update_posterior: bool = True,
 ) -> FeedbackResult:
     """Apply one feedback event to one belief.
 
@@ -92,8 +93,12 @@ def apply_feedback(
     2. Reject zero valence: a no-update event is not a successful update,
        and pre-commit #5 says feedback_history records every successful
        update — so a zero call has no row to write.
-    3. Bayesian-update alpha or beta by valence sign.
-    4. Persist the new posterior on the belief row.
+    3. Bayesian-update alpha or beta by valence sign — UNLESS
+       `update_posterior` is False, in which case the posterior is left
+       untouched (audit-only; #1086). Retrieval-exposure records an event
+       for the recurrence axis without being treated as truth-evidence.
+    4. Persist the new posterior on the belief row (skipped when
+       `update_posterior` is False — nothing changed).
     5. Append one row to feedback_history (created_at = `now` or UTC now).
     6. Propagate the signal through outbound edges (#1058): each
        attenuated delta from `store.propagate_valence` is applied via a
@@ -124,11 +129,18 @@ def apply_feedback(
 
     prior_alpha: float = b.alpha
     prior_beta: float = b.beta
-    new_alpha, new_beta = _bayesian_update(b, valence)
-
-    b.alpha = new_alpha
-    b.beta = new_beta
-    store.update_belief(b)
+    if update_posterior:
+        new_alpha, new_beta = _bayesian_update(b, valence)
+        b.alpha = new_alpha
+        b.beta = new_beta
+        store.update_belief(b)
+    else:
+        # Audit-only (#1086): record the event so exposure frequency stays
+        # recoverable (the recurrence axis), but do NOT move the posterior.
+        # A retrieval is exposure, not endorsement; counting every surfacing
+        # as positive evidence inflated whatever recurs. The posterior is
+        # left exactly as-is; the feedback_history row is still written.
+        new_alpha, new_beta = prior_alpha, prior_beta
 
     timestamp: str = now if now is not None else _utc_now_iso()
     event_id: int = store.insert_feedback_event(
@@ -149,7 +161,7 @@ def apply_feedback(
         source=source,
     )
 
-    if propagate and _propagation_enabled():
+    if update_posterior and propagate and _propagation_enabled():
         deltas = store.propagate_valence(belief_id, valence)
         # Sorted for a deterministic feedback_history row order
         # regardless of edge-iteration order inside the BFS.
diff --git a/src/aelfrice/hook_search.py b/src/aelfrice/hook_search.py
index 7b544f174..7ebb4a0a8 100644
--- a/src/aelfrice/hook_search.py
+++ b/src/aelfrice/hook_search.py
@@ -12,11 +12,14 @@
    to the agent". The README's Bayesian-memory claim depends on this
    loop closing.
 
-2. Hook-driven valence is implicit and weaker than explicit user
-   feedback. A retrieval is exposure, not endorsement. Use a small
-   positive valence (`HOOK_RETRIEVAL_VALENCE = 0.1`) so a thousand
-   retrievals over a few months don't dominate the posterior the way a
-   handful of explicit user thumbs-up should.
+2. A retrieval is exposure, not endorsement. Since #1086 the hook
+   records the exposure event to `feedback_history` (so surfacing
+   frequency stays recoverable) but does NOT move the Bayesian posterior
+   by default: counting every surfacing as positive evidence let whatever
+   recurs float above genuine knowledge. The legacy behaviour — a small
+   positive `HOOK_RETRIEVAL_VALENCE = 0.1` per retrieval — is restored by
+   setting `AELFRICE_EXPOSURE_UPDATES_POSTERIOR=1` (benchmark A/B and
+   rollback).
 
 The module exposes two functions: `search_for_prompt`, the hook's
 top-level call (retrieve + record), and `record_retrieval`, the audit
@@ -31,6 +34,7 @@
 """
 from __future__ import annotations
 
+import os
 import sys
 import traceback
 from typing import IO, Final, Iterable
@@ -45,6 +49,23 @@
 retrieval row. ARCHITECTURE.md commits this string publicly; downstream
 analysis (e.g. `SELECT ... WHERE source = 'hook'`) depends on it."""
 
+ENV_EXPOSURE_UPDATES_POSTERIOR: Final[str] = "AELFRICE_EXPOSURE_UPDATES_POSTERIOR"
+"""Opt-in to the legacy behaviour where a hook retrieval moves the
+Bayesian posterior. Default is OFF (#1086): a retrieval is exposure, not
+endorsement, and counting every surfacing as positive evidence inflated
+whatever recurs — measured on a real store, junk (session scaffolding,
+fragments) accumulated MORE exposure than genuine knowledge and floated
+above it. With the flag unset, hook retrievals are recorded to
+feedback_history (so exposure frequency stays recoverable) but leave the
+posterior untouched. Set to "1" to restore the pre-#1086 posterior update
+(kept for benchmark A/B and rollback). Read per call so a flip is honoured
+without restart."""
+
+
+def _exposure_updates_posterior() -> bool:
+    return os.environ.get(ENV_EXPOSURE_UPDATES_POSTERIOR, "0") == "1"
+
+
 HOOK_RETRIEVAL_VALENCE: Final[float] = 0.1
 """Per-belief positive valence written for each hook-driven retrieval.
 
@@ -99,6 +120,7 @@ def record_retrieval(
     needing to inspect the audit log.
     """
     serr: IO[str] = stderr if stderr is not None else sys.stderr
+    update_posterior: bool = _exposure_updates_posterior()
     written: int = 0
     stamped_ids: list[str] = []
     for b in beliefs:
@@ -108,6 +130,7 @@ def record_retrieval(
                 b.id,
                 valence,
                 source,
+                update_posterior=update_posterior,
             )
             written += 1
             stamped_ids.append(b.id)
diff --git a/tests/regression/test_hook_to_feedback_history_end_to_end.py b/tests/regression/test_hook_to_feedback_history_end_to_end.py
index 965847443..1210ef796 100644
--- a/tests/regression/test_hook_to_feedback_history_end_to_end.py
+++ b/tests/regression/test_hook_to_feedback_history_end_to_end.py
@@ -99,9 +99,41 @@ def test_hook_fire_writes_feedback_rows_tagged_hook(
     assert {e.belief_id for e in hook_events} >= {"F1", "F2"}
 
 
-def test_hook_fire_increments_alpha_not_beta(
+def test_hook_fire_records_exposure_without_moving_posterior(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None:
+    """#1086: firing the hook records the retrieval to feedback_history
+    (recurrence axis) but leaves the posterior untouched by default —
+    exposure is not endorsement."""
+    db = tmp_path / "memory.db"
+    s = MemoryStore(str(db))
+    s.insert_belief(_mk("F1", "the kitchen is full of bananas"))
+    s.close()
+    _set_db(monkeypatch, db)
+
+    user_prompt_submit(
+        stdin=io.StringIO(_payload("are there bananas in the kitchen")),
+        stdout=io.StringIO(),
+    )
+
+    s2 = MemoryStore(str(db))
+    try:
+        b = s2.get_belief("F1")
+        events = s2.count_feedback_events()
+    finally:
+        s2.close()
+    assert b is not None
+    assert b.alpha == 1.0  # unchanged — audit-only
+    assert b.beta == 1.0
+    assert events >= 1  # exposure still recorded
+
+
+def test_hook_fire_legacy_flag_increments_alpha(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """AELFRICE_EXPOSURE_UPDATES_POSTERIOR=1 restores the pre-#1086
+    behaviour: a hook fire bumps alpha by HOOK_RETRIEVAL_VALENCE."""
+    monkeypatch.setenv("AELFRICE_EXPOSURE_UPDATES_POSTERIOR", "1")
     db = tmp_path / "memory.db"
     s = MemoryStore(str(db))
     s.insert_belief(_mk("F1", "the kitchen is full of bananas"))
diff --git a/tests/test_feedback_apply.py b/tests/test_feedback_apply.py
index 9f0182887..d0271e34d 100644
--- a/tests/test_feedback_apply.py
+++ b/tests/test_feedback_apply.py
@@ -214,3 +214,35 @@ def test_unknown_belief_id_does_not_write_history() -> None:
     with pytest.raises(ValueError):
         apply_feedback(s, "nonexistent", valence=1.0, source="user")
     assert s.count_feedback_events() == 0
+
+
+# --- Audit-only exposure (update_posterior=False, #1086) -----------------
+
+
+def test_update_posterior_false_leaves_alpha_beta_unchanged() -> None:
+    """A retrieval is exposure, not endorsement: audit-only feedback must
+    not move the Bayesian posterior."""
+    s = _store_with(_mk(alpha=2.0, beta=3.0))
+    apply_feedback(s, "b1", valence=0.1, source="hook", update_posterior=False)
+    got = s.get_belief("b1")
+    assert got is not None
+    assert (got.alpha, got.beta) == (2.0, 3.0)
+
+
+def test_update_posterior_false_still_writes_audit_row() -> None:
+    """Exposure is recorded to feedback_history so surfacing frequency
+    stays recoverable for the recurrence axis."""
+    s = _store_with(_mk())
+    apply_feedback(s, "b1", valence=0.1, source="hook", update_posterior=False)
+    assert s.count_feedback_events() == 1
+
+
+def test_update_posterior_false_result_reports_no_change() -> None:
+    res = apply_feedback(
+        s := _store_with(_mk(alpha=2.0, beta=3.0)),
+        "b1", valence=0.1, source="hook", update_posterior=False,
+    )
+    assert res.new_alpha == res.prior_alpha == 2.0
+    assert res.new_beta == res.prior_beta == 3.0
+    assert res.propagated == []  # propagation is a posterior effect — skipped
+    s.close()
diff --git a/tests/test_hook_search.py b/tests/test_hook_search.py
index 58888fcb7..c7cf2fe48 100644
--- a/tests/test_hook_search.py
+++ b/tests/test_hook_search.py
@@ -101,7 +101,27 @@ def test_record_retrieval_empty_iterable_writes_no_rows() -> None:
     assert s.count_feedback_events() == 0
 
 
-def test_record_retrieval_updates_alpha_not_beta() -> None:
+def test_record_retrieval_audit_only_does_not_move_posterior() -> None:
+    """#1086: a hook retrieval is exposure, not endorsement. By default it
+    records an audit row for the recurrence axis but leaves the Bayesian
+    posterior untouched — counting every surfacing as evidence let whatever
+    recurs float above genuine knowledge."""
+    s = _seed(_mk("b1"))
+    written = record_retrieval(s, [s.get_belief("b1")])  # type: ignore[list-item]
+    assert written == 1
+    b = s.get_belief("b1")
+    assert b is not None
+    assert b.alpha == 1.0  # unchanged
+    assert b.beta == 1.0
+    assert s.count_feedback_events() == 1  # audit row still written
+
+
+def test_record_retrieval_legacy_flag_updates_alpha(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """AELFRICE_EXPOSURE_UPDATES_POSTERIOR=1 restores the pre-#1086
+    behaviour: a hook retrieval bumps alpha by HOOK_RETRIEVAL_VALENCE."""
+    monkeypatch.setenv("AELFRICE_EXPOSURE_UPDATES_POSTERIOR", "1")
     s = _seed(_mk("b1"))
     record_retrieval(s, [s.get_belief("b1")])  # type: ignore[list-item]
     b = s.get_belief("b1")