diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index 68c719a12..54e221a4d 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -26,7 +26,7 @@
 
 ### Fixed
 
-- Stopped Mnemopi retention from overflowing the embedding model's context window: `embed()` now caps each input at `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS` (default 32000 chars) so a long multi-turn `MnemopiSessionState.retainMessages` transcript can't make llama.cpp's `/embeddings` server reject the request with `request (N tokens) exceeds the available context size` and silently drop vector recall for that memory ([#3126](https://github.com/can1357/oh-my-pi/issues/3126)).
+- Stopped Mnemopi retention from overflowing the embedding model's context window: `embed()` now caps each input at `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS` (default 8192 chars) so a long multi-turn `MnemopiSessionState.retainMessages` transcript can't make llama.cpp's `/embeddings` server reject the request with `request (N tokens) exceeds the available context size` and silently drop vector recall for that memory ([#3126](https://github.com/can1357/oh-my-pi/issues/3126)).
 
 ## [16.1.7] - 2026-06-20
 
diff --git a/packages/mnemopi/CHANGELOG.md b/packages/mnemopi/CHANGELOG.md
index dfb11763d..ba3a5658c 100644
--- a/packages/mnemopi/CHANGELOG.md
+++ b/packages/mnemopi/CHANGELOG.md
@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Capped per-input length in `embed()` at `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS` (default 32000 chars, override via the env var or `embeddings.maxInputChars` runtime option; `0` disables) so a long retention transcript can no longer overflow the embedding model's context window. llama.cpp's `/embeddings` server used to reject the request with `request (N tokens) exceeds the available context size`, silently dropping vector recall for that memory ([#3126](https://github.com/can1357/oh-my-pi/issues/3126)).
+- Capped per-input length in `embed()` at `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS` (default 8192 chars, override via the env var or `embeddings.maxInputChars` runtime option; `0` disables) so a long retention transcript can no longer overflow the embedding model's context window. llama.cpp's `/embeddings` server used to reject the request with `request (N tokens) exceeds the available context size`, silently dropping vector recall for that memory ([#3126](https://github.com/can1357/oh-my-pi/issues/3126)).
 
 ## [16.1.3] - 2026-06-19
 
diff --git a/packages/mnemopi/src/config.ts b/packages/mnemopi/src/config.ts
index 3bee5a23c..56ee807ca 100644
--- a/packages/mnemopi/src/config.ts
+++ b/packages/mnemopi/src/config.ts
@@ -110,13 +110,13 @@ export function embeddingsDisabled(env: Env = process.env): boolean {
  * source gives both backends deterministic behavior and prevents the silent
  * recall degradation we saw in issue #3126.
  *
- * Default `32000` ≈ 8k English tokens (4 chars/token) or ~16k–32k CJK tokens —
- * fits 8k-context models (bge-m3, text-embedding-3) for English and most CJK
- * content. Lower it for 512-token models like `BAAI/bge-small-en-v1.5`, raise
- * it for larger contexts. `0` disables the cap.
+ * Default `8192` chars is intentionally conservative for 8192-token embedding
+ * contexts (bge-m3, OpenAI text-embedding-3) and CJK-heavy transcripts. Raise
+ * it for larger local contexts (for example Qwen3-Embedding with 32k ctx).
+ * `0` disables the cap.
  */
 export function embeddingMaxInputChars(env: Env = process.env): number {
-	return Math.max(0, envInt("MNEMOPI_EMBEDDING_MAX_INPUT_CHARS", 32000, env));
+	return Math.max(0, envInt("MNEMOPI_EMBEDDING_MAX_INPUT_CHARS", 8192, env));
 }
 
 export function isApiEmbeddingModel(model = embeddingModel(), env: Env = process.env): boolean {
diff --git a/packages/mnemopi/src/core/embeddings.ts b/packages/mnemopi/src/core/embeddings.ts
index 37f79b10c..2d84ca7a1 100644
--- a/packages/mnemopi/src/core/embeddings.ts
+++ b/packages/mnemopi/src/core/embeddings.ts
@@ -124,7 +124,7 @@ export function embeddingsDisabled(): boolean {
  * Resolved per-input character cap for {@link embed}.
  *
  * Reads (in order): the active runtime scope's `embeddings.maxInputChars`, then
- * `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS`, then the bundled `32000` default. `0`
+ * `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS`, then the bundled `8192` default. `0`
  * disables the cap entirely.
  */
 function effectiveMaxInputChars(): number {
@@ -132,7 +132,7 @@ function effectiveMaxInputChars(): number {
 	if (override !== undefined) return Math.max(0, Math.trunc(override));
 	const envValue = Number.parseInt($env.MNEMOPI_EMBEDDING_MAX_INPUT_CHARS ?? "", 10);
 	if (Number.isFinite(envValue) && envValue >= 0) return envValue;
-	return 32000;
+	return 8192;
 }
 
 /**
diff --git a/packages/mnemopi/test/embedding-input-cap.test.ts b/packages/mnemopi/test/embedding-input-cap.test.ts
index 050e729be..46217ac3c 100644
--- a/packages/mnemopi/test/embedding-input-cap.test.ts
+++ b/packages/mnemopi/test/embedding-input-cap.test.ts
@@ -13,7 +13,7 @@ import { withMnemopiRuntimeOptions } from "@oh-my-pi/pi-mnemopi/core/runtime-opt
  * overflow whatever ctx the embedding server was started with — llama.cpp
  * rejects oversized requests with `request (N tokens) exceeds the available
  * context size`, OpenAI silently right-truncates. `embed()` now caps each
- * input to `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS` (default 32000) before the
+ * input to `MNEMOPI_EMBEDDING_MAX_INPUT_CHARS` (default 8192) before the
  * provider sees it.
  */
 function captureProvider(): {
@@ -57,7 +57,7 @@ describe("embed() input cap (#3126)", () => {
 		expect(provider.calls).toHaveLength(1);
 		const [seenShort, seenHuge] = provider.calls[0] ?? [];
 		expect(seenShort).toBe("short");
-		expect(seenHuge?.length).toBe(32_000);
+		expect(seenHuge?.length).toBe(8192);
 	});
 
 	it("honors MNEMOPI_EMBEDDING_MAX_INPUT_CHARS env override", async () => {