diff --git a/src/responses/state.ts b/src/responses/state.ts
index 3345e4626..8a7e4963c 100644
--- a/src/responses/state.ts
+++ b/src/responses/state.ts
@@ -27,7 +27,7 @@ interface StoredResponseState {
   /** v1 Cursor-only metadata, accepted only while loading old snapshots. */
   conversationId?: string;
   cursorCheckpointUsable?: boolean;
-  /** Approximate in-memory size, computed locally at insert time (never trusted from disk). */
+  /** Serialized UTF-8 entry size, computed locally at insert time (never trusted from disk). */
   sizeBytes?: number;
 }
 
@@ -49,21 +49,28 @@ export function getStoredResponseBytesForTests(): number {
   return storedResponseBytes;
 }
 
-/** The ONLY size computation: approximate entry weight from its items payload. */
-function measuredEntry(entry: Omit<StoredResponseState, "sizeBytes">): StoredResponseState {
-  let sizeBytes = 0;
+/** The ONLY size computation: serialized UTF-8 bytes retained by the map entry. */
+function measuredEntry(id: string, entry: Omit<StoredResponseState, "sizeBytes">): StoredResponseState | null {
   try {
-    sizeBytes = JSON.stringify(entry.items).length;
+    const serialized = JSON.stringify([id, entry]);
+    return { ...entry, sizeBytes: Buffer.byteLength(serialized, "utf8") };
   } catch {
-    /* unserializable items: weightless rather than fatal */
+    // A value whose retained size cannot be measured cannot safely enter a byte-bounded store.
+    return null;
   }
-  return { ...entry, sizeBytes };
 }
 
-/** The ONLY insertion point: keeps the byte counter consistent on replacement. */
+/**
+ * The ONLY insertion point: measure before mutating, reject an over-cap candidate without
+ * evicting unrelated valid chains, and remove a stale same-id value before rejecting it.
+ */
 function setEntry(id: string, entry: Omit<StoredResponseState, "sizeBytes">): void {
+  const measured = measuredEntry(id, entry);
+  if (!measured || (measured.sizeBytes ?? 0) > byteCap()) {
+    deleteEntry(id);
+    return;
+  }
   deleteEntry(id);
-  const measured = measuredEntry(entry);
   storedResponseBytes += measured.sizeBytes ?? 0;
   states.set(id, measured);
 }
@@ -326,7 +333,7 @@ function pruneResponses(at = now()): void {
     deleteEntry(oldest);
   }
   // Byte high-water eviction, oldest-first (Map preserves insertion order).
-  while (storedResponseBytes > byteCap() && states.size > 1) {
+  while (storedResponseBytes > byteCap() && states.size > 0) {
     const oldest = states.keys().next().value;
     if (!oldest) break;
     deleteEntry(oldest);
@@ -371,7 +378,7 @@ export function previousResponseProviderState(responseId: string | undefined): O
 export interface ResponseStateMetrics {
   /** Number of retained continuation entries currently in RAM. */
   count: number;
-  /** Sum of serialized item bytes across all entries (the running in-memory byte accounting).
+  /** Sum of serialized map-entry bytes across all entries (the running in-memory byte accounting).
    * Because expanded previous_response_id chains share prefix item references, this is an UPPER
    * bound on true heap (it re-counts shared history), never an under-count — safe as a
    * memory-pressure signal. */
diff --git a/structure/02_config-and-codex-home.md b/structure/02_config-and-codex-home.md
index af519117f..6bfd28cc4 100644
--- a/structure/02_config-and-codex-home.md
+++ b/structure/02_config-and-codex-home.md
@@ -57,6 +57,18 @@ are consumed incrementally and at most 512 stale files are attempted per process
 - 다른 대안 대신 이 방식을 선택한 이유: It repairs known remnants without broad authority over unrelated temp files or active writers.
 - 장점, 단점 및 영향: Old dead-PID files are reclaimed automatically; locked or conservatively classified files remain for a later retry.
 
+In-memory continuation entries are measured in serialized UTF-8 bytes before admission. An entry
+larger than the 64 MiB store budget is not retained, and replacing an existing response ID with an
+oversized value removes the stale value without evicting unrelated valid chains.
+
+[Decision Log]
+- 목적과 의도: Make the Responses continuation byte budget a hard invariant even for one image-heavy or malformed entry.
+- 기존 구현 및 제약 조건: Oldest-first pruning stopped when one entry remained, so a single value could exceed the nominal 64 MiB cap; continuation misses are recoverable but unbounded retention is not.
+- 검토한 주요 대안: Keep the newest entry regardless of size; evict all other entries before admitting it; reject the candidate atomically.
+- 선택한 방식: Measure UTF-8 bytes before mutation, reject an over-budget candidate, and remove only a stale entry with the same response ID.
+- 다른 대안 대신 이 방식을 선택한 이유: Evicting unrelated chains cannot make an individually oversized value fit, while retaining it breaks the advertised memory bound.
+- 장점, 단점 및 영향: RAM accounting remains at or below the cap; an oversized continuation becomes a normal previous-response miss and may require the client to resend history.
+
 ## Config surface
 
 `src/types.ts` is the shape and `src/config.ts` is the loader; neither is reproduced here. What
diff --git a/tests/responses-state.test.ts b/tests/responses-state.test.ts
index 1e18efda2..f336b4862 100644
--- a/tests/responses-state.test.ts
+++ b/tests/responses-state.test.ts
@@ -554,6 +554,94 @@ describe("Responses previous_response_id state", () => {
     }
   });
 
+  test("a single over-cap entry is rejected without evicting an unrelated valid chain", () => {
+    setResponseStateByteCapForTests(1_000);
+    try {
+      const small = buildResponseJSON([
+        { type: "text_delta", text: "ok" },
+        { type: "done" },
+      ], "gpt-5.5");
+      rememberResponseState({ model: "gpt-5.5", input: "small" }, small);
+      const smallBytes = getStoredResponseBytesForTests();
+
+      const oversized = buildResponseJSON([
+        { type: "text_delta", text: "x".repeat(3_000) },
+        { type: "done" },
+      ], "gpt-5.5");
+      rememberResponseState({ model: "gpt-5.5", input: "too large" }, oversized);
+
+      expect(expandPreviousResponseInput({
+        model: "gpt-5.5", previous_response_id: oversized.id, input: "next",
+      })).toEqual({ model: "gpt-5.5", previous_response_id: oversized.id, input: "next" });
+      expect((expandPreviousResponseInput({
+        model: "gpt-5.5", previous_response_id: small.id, input: "next",
+      }) as { input: unknown[] }).input).toHaveLength(3);
+      expect(getStoredResponseBytesForTests()).toBe(smallBytes);
+      expect(responseStateMetrics().totalBytes).toBeLessThanOrEqual(1_000);
+    } finally {
+      setResponseStateByteCapForTests(null);
+    }
+  });
+
+  test("an over-cap replacement removes stale state for the same response id", () => {
+    setResponseStateByteCapForTests(1_000);
+    try {
+      const responseId = "resp_replaced_over_cap";
+      rememberResponseState(
+        { model: "gpt-5.5", input: "small" },
+        { id: responseId, status: "completed", output: [{ type: "message", content: "ok" }] },
+      );
+      expect(responseStateMetrics().count).toBe(1);
+
+      rememberResponseState(
+        { model: "gpt-5.5", input: "replacement" },
+        { id: responseId, status: "completed", output: [{ type: "message", content: "x".repeat(3_000) }] },
+      );
+
+      const next = { model: "gpt-5.5", previous_response_id: responseId, input: "next" };
+      expect(expandPreviousResponseInput(next)).toEqual(next);
+      expect(responseStateMetrics()).toEqual({ count: 0, totalBytes: 0, largestBytes: 0, oldestAgeMs: 0 });
+    } finally {
+      setResponseStateByteCapForTests(null);
+    }
+  });
+
+  test("byte admission counts UTF-8 bytes instead of JavaScript code units", () => {
+    setResponseStateByteCapForTests(1_000);
+    try {
+      const responseId = "resp_multibyte_over_cap";
+      rememberResponseState(
+        { model: "gpt-5.5", input: "한".repeat(400) },
+        { id: responseId, status: "completed", output: [] },
+      );
+
+      const next = { model: "gpt-5.5", previous_response_id: responseId, input: "next" };
+      expect(expandPreviousResponseInput(next)).toEqual(next);
+      expect(getStoredResponseBytesForTests()).toBe(0);
+    } finally {
+      setResponseStateByteCapForTests(null);
+    }
+  });
+
+  test("snapshot loading rejects a single entry above the active byte cap", () => {
+    setResponseStateByteCapForTests(1_000);
+    try {
+      mkdirSync(home, { recursive: true });
+      writeFileSync(join(home, "responses-state.json"), JSON.stringify({
+        version: 2,
+        states: [["resp_snapshot_over_cap", {
+          createdAt: Date.now(),
+          items: [{ role: "user", content: "x".repeat(3_000) }],
+        }]],
+      }));
+
+      expect(previousResponseProviderState("resp_snapshot_over_cap")).toBeUndefined();
+      expect(responseStateMetrics()).toEqual({ count: 0, totalBytes: 0, largestBytes: 0, oldestAgeMs: 0 });
+    } finally {
+      setResponseStateByteCapForTests(null);
+    }
+  });
+
   test("byte accounting survives restart (sizes recomputed on load)", async () => {
     setResponseStateByteCapForTests(4_000);
     try {
@@ -609,7 +697,9 @@ describe("Responses previous_response_id state", () => {
       for (let i = 0; i < 1_050; i++) {
         const body = { model: "cursor/grok-4.5", input: `${bulk}-000${String(i % 10)}`, store: false };
         const json = buildResponseJSON([{ type: "text_delta", text: "ok" }, { type: "done" }], "cursor/grok-4.5");
-        rememberResponseState(body, json, { cursor: { conversationId: `conv_c${i}` } }, { force: true });
+        rememberResponseState(body, json, {
+          cursor: { conversationId: `conv_c${String(i).padStart(4, "0")}` },
+        }, { force: true });
         lastId = json.id as string;
         if (perEntryBytes === 0) perEntryBytes = getStoredResponseBytesForTests();
       }