diff --git a/__tests__/unit/cat/forget-word-boundary.test.ts b/__tests__/unit/cat/forget-word-boundary.test.ts
new file mode 100644
index 000000000..83aeaf39e
--- /dev/null
+++ b/__tests__/unit/cat/forget-word-boundary.test.ts
@@ -0,0 +1,117 @@
+/**
+ * A forget phrase must match WORDS, not letter runs.
+ *
+ * The containment branch compared raw substrings both ways, and
+ * MIN_FORGET_FRAGMENT_CHARS allows four-character facts. So "work" was
+ * contained in "network", "framework", "coworking" and "homework — every one
+ * of those memories was deleted by a user asking to forget "work".
+ *
+ * This is the opposite failure to #563 finding 8, and the worse one: that bug
+ * told the user something was still there when it was gone, this one silently
+ * destroys memories they never asked to remove, and Cat then reports them as
+ * deleted — accurately, which is what makes it hard to notice.
+ *
+ * bitbaum/orangecat#563 finding 9.
+ */
+
+import { forgetMemoriesMatching } from '@/services/cat/memory';
+import type { AnySupabaseClient } from '@/lib/supabase/types';
+
+interface Row {
+  id: string;
+  content: string;
+}
+
+function supabaseWithCorpus(corpus: Row[]) {
+  const deleted: string[][] = [];
+  const client = {
+    from: () => ({
+      select: () => ({ eq: async () => ({ data: corpus, error: null }) }),
+      delete: () => ({
+        eq: () => ({
+          in: async (_col: string, ids: string[]) => {
+            deleted.push(ids);
+            return { error: null };
+          },
+        }),
+      }),
+    }),
+    rpc: async () => ({ data: [], error: null }),
+  } as unknown as AnySupabaseClient;
+  return { client, deleted };
+}
+
+const CORPUS: Row[] = [
+  { id: 'm1', content: 'Builds neural networks for a living' },
+  { id: 'm2', content: 'Prefers the React framework' },
+  { id: 'm3', content: 'Uses a coworking space in Zurich' },
+  { id: 'm4', content: 'Does not like work on weekends' },
+  { id: 'm5', content: 'Knows French' },
+];
+
+describe('forget matches words, not letter runs', () => {
+  it('forgetting "work" leaves network, framework and coworking alone', async () => {
+    const { client } = supabaseWithCorpus(CORPUS);
+    const result = await forgetMemoriesMatching(client, 'u1', ['work']);
+
+    // Only the memory that contains the actual WORD "work".
+    expect(result.deleted).toEqual(['Does not like work on weekends']);
+    expect(result.deleted).not.toContain('Builds neural networks for a living');
+    expect(result.deleted).not.toContain('Prefers the React framework');
+    expect(result.deleted).not.toContain('Uses a coworking space in Zurich');
+  });
+
+  it('still matches the whole word inside a longer sentence', async () => {
+    // The fix must not break the thing containment is FOR.
+    const { client } = supabaseWithCorpus(CORPUS);
+    const result = await forgetMemoriesMatching(client, 'u1', ['French']);
+
+    expect(result.deleted).toEqual(['Knows French']);
+  });
+
+  it('matches a multi-word phrase on word boundaries', async () => {
+    const { client } = supabaseWithCorpus([
+      { id: 'a', content: 'Uses a coworking space in Zurich' },
+      { id: 'b', content: 'Has a coworkings pass' },
+    ]);
+    const result = await forgetMemoriesMatching(client, 'u1', ['coworking space']);
+
+    expect(result.deleted).toEqual(['Uses a coworking space in Zurich']);
+  });
+
+  it('does not match a prefix of a longer word', async () => {
+    const { client } = supabaseWithCorpus([
+      { id: 'a', content: 'Is a constructor at heart' },
+      { id: 'b', content: 'Has one constraint: weekends' },
+    ]);
+    // "constra" is a fragment of "constraint" — not a word anywhere.
+    const result = await forgetMemoriesMatching(client, 'u1', ['constru']);
+
+    expect(result.deleted).toEqual([]);
+    expect(result.notFound).toEqual(['constru']);
+  });
+
+  it('treats an accented word as a whole word', async () => {
+    // `\b` is ASCII-only in JS — it would see "café" as ending after "caf" and
+    // happily match inside "cafétéria". The stemmer does NOT unify these (its
+    // suffix list cannot turn "cafétéria" into "café"), so containment is the
+    // only branch in play and this isolates the boundary check.
+    const { client } = supabaseWithCorpus([
+      { id: 'a', content: 'Runs a café in Zurich' },
+      { id: 'b', content: 'Prefers cafétéria food' },
+    ]);
+    const result = await forgetMemoriesMatching(client, 'u1', ['café']);
+
+    expect(result.deleted).toEqual(['Runs a café in Zurich']);
+  });
+
+  it('still lets the STEMMER unify a plural — that is its job, not a boundary bug', async () => {
+    // "cafés" → "café" is the same mechanism that makes "weekends" match
+    // "weekend" and "photography" match "photographer". Deliberately unchanged:
+    // this fix narrows raw containment, it does not touch stemming.
+    const { client } = supabaseWithCorpus([{ id: 'a', content: 'Reviews cafés for a blog' }]);
+    const result = await forgetMemoriesMatching(client, 'u1', ['café']);
+
+    expect(result.deleted).toEqual(['Reviews cafés for a blog']);
+  });
+});
diff --git a/src/services/cat/memory.ts b/src/services/cat/memory.ts
index 003699933..d4e75f95b 100644
--- a/src/services/cat/memory.ts
+++ b/src/services/cat/memory.ts
@@ -150,6 +150,41 @@ function stemWord(word: string): string {
   return s;
 }
 
+/**
+ * Does `haystack` contain `needle` as WHOLE WORDS?
+ *
+ * The containment branch used raw `String.includes` in both directions, and
+ * MIN_FORGET_FRAGMENT_CHARS lets a four-character fact through. So "work" was
+ * contained in "network", "framework", "coworking" and "homework": asking Cat
+ * to forget "work" deleted every one of those memories, and Cat then reported
+ * them as removed — accurately, which is exactly what made it hard to notice.
+ *
+ * Boundaries are checked by CHARACTER CLASS rather than a `\b` regex, because
+ * `\b` is ASCII-only in JavaScript: it treats "café" as ending after "caf",
+ * so "café" would match inside "cafés" while plain words behaved correctly.
+ * `\p{L}` and `\p{N}` cover the accented alphabet the tokenizer above already
+ * speaks. bitbaum/orangecat#563 finding 9.
+ */
+const WORD_CHAR = /[\p{L}\p{N}]/u;
+export function containsWholeWords(haystack: string, needle: string): boolean {
+  if (!needle || !haystack) {
+    return false;
+  }
+  for (let from = 0; from <= haystack.length - needle.length; ) {
+    const at = haystack.indexOf(needle, from);
+    if (at === -1) {
+      return false;
+    }
+    const before = at === 0 ? '' : haystack[at - 1]!;
+    const after = haystack[at + needle.length] ?? '';
+    if (!WORD_CHAR.test(before) && !WORD_CHAR.test(after)) {
+      return true;
+    }
+    from = at + 1;
+  }
+  return false;
+}
+
 /** Significant stemmed words of a phrase (short glue words dropped). */
 function significantStems(text: string): Set<string> {
   return new Set(
@@ -257,10 +292,17 @@ export async function forgetMemoriesMatching(
     for (let i = 0; i < corpus.length; i++) {
       const m = corpus[i];
       const c = m.content.toLowerCase();
-      // Containment either way ("photography" ⊂ "Has photography skills…"),
-      // or enough shared stems (see requiredStemHits — two-word facts need
-      // TWO hits, so a single shared "skills" can't delete a stranger).
-      if (c.includes(norm) || norm.includes(c) || stemOverlapMatches(factStems, memoryStems[i])) {
+      // Containment either way, on WORD boundaries ("photography" ⊂ "Has
+      // photography skills…"), or enough shared stems (see requiredStemHits —
+      // two-word facts need TWO hits, so one shared "skills" cannot delete a
+      // stranger).
+      // Whole words only: raw containment let a 4-char fact delete every memory
+      // that merely SPELLED it ("work" ⊂ network/framework/coworking).
+      if (
+        containsWholeWords(c, norm) ||
+        containsWholeWords(norm, c) ||
+        stemOverlapMatches(factStems, memoryStems[i])
+      ) {
         doomed.set(m.id, m.content);
         matched = true;
       }