diff --git a/web/src/components/MemoEditor/Editor/tagAutocomplete.ts b/web/src/components/MemoEditor/Editor/tagAutocomplete.ts
index 0574ac37..73c0d36a 100644
--- a/web/src/components/MemoEditor/Editor/tagAutocomplete.ts
+++ b/web/src/components/MemoEditor/Editor/tagAutocomplete.ts
@@ -3,6 +3,23 @@ import type { Extension } from "@codemirror/state";
 import { isTagIntroducerAt } from "@/utils/tag-grammar";
 import { isLiteralTagPosition, tagMatchBefore } from "./markdownTagRanges";
 
+/**
+ * Ranks a candidate tag against the typed text (both already lower-cased).
+ * Lower sorts first; `undefined` means the candidate does not match.
+ *
+ * Nested tags are paths, so a match inside the path counts: typing `mem` has to
+ * reach `software/hosted/Memos`, which a full-path prefix test cannot do. The
+ * tiers keep that from turning into noise — a segment start outranks an
+ * incidental hit in the middle of a word.
+ */
+const matchRank = (tag: string, typed: string): number | undefined => {
+  const index = tag.indexOf(typed);
+  if (index < 0) return undefined;
+  if (index === 0) return 0; // Full-path prefix.
+  if (tag[index - 1] === "/") return 1; // Path-segment start.
+  return 2; // Anywhere else.
+};
+
 export function makeTagCompletionSource(getTags: () => string[]) {
   return (ctx: CompletionContext): CompletionResult | null => {
     const match = tagMatchBefore(ctx.state, ctx.pos);
@@ -13,10 +30,15 @@ export function makeTagCompletionSource(getTags: () => string[]) {
 
     const typed = (match?.value ?? "").toLowerCase();
     const options = getTags()
-      .filter((tag) => tag.toLowerCase().startsWith(typed))
-      .map((tag) => ({ label: tag, type: "keyword" }));
+      .map((tag) => ({ tag, rank: matchRank(tag.toLowerCase(), typed) }))
+      .filter((candidate): candidate is { tag: string; rank: number } => candidate.rank !== undefined)
+      // Stable sort, so tags keep their incoming order within a tier.
+      .sort((a, b) => a.rank - b.rank)
+      .map(({ tag }) => ({ label: tag, type: "keyword" }));
     if (options.length === 0) return null;
-    return { from: match ? match.from + 1 : ctx.pos, options };
+    // `filter: false` keeps this ranking: CodeMirror would otherwise re-filter
+    // and re-score the options with its own fuzzy matcher.
+    return { from: match ? match.from + 1 : ctx.pos, options, filter: false };
   };
 }
 
diff --git a/web/tests/editor-tag-autocomplete.test.ts b/web/tests/editor-tag-autocomplete.test.ts
index 16e36ef3..a801c546 100644
--- a/web/tests/editor-tag-autocomplete.test.ts
+++ b/web/tests/editor-tag-autocomplete.test.ts
@@ -17,6 +17,19 @@ describe("tag autocomplete", () => {
     expect(result?.options.map((o) => o.label)).toEqual(["todo", "today"]);
   });
 
+  it("matches a segment of a nested tag path", () => {
+    expect(complete("hello #Mem", 10, ["software/hosted/Memos"])?.options.map((o) => o.label)).toEqual(["software/hosted/Memos"]);
+  });
+
+  it("ranks full-path prefixes above segment starts above loose substrings", () => {
+    const result = complete("#work", 5, ["home/paperwork", "team/work-log", "work/project"]);
+    expect(result?.options.map((o) => o.label)).toEqual(["work/project", "team/work-log", "home/paperwork"]);
+  });
+
+  it("keeps its own ranking instead of CodeMirror's fuzzy filter", () => {
+    expect(complete("#work", 5, ["work"])?.filter).toBe(false);
+  });
+
   it("returns null on a bare # with nothing typed", () => {
     expect(complete("hello #", 7, ["todo"])).toBeNull();
   });