diff --git a/src/tangara/database/database.cpp b/src/tangara/database/database.cpp
index 67893b6e..9854f693 100644
--- a/src/tangara/database/database.cpp
+++ b/src/tangara/database/database.cpp
@@ -7,6 +7,7 @@
 #include "database/database.hpp"
 
 #include <bits/ranges_algo.h>
+#include <stdint.h>
 #include <algorithm>
 #include <cctype>
 #include <cstdint>
@@ -21,6 +22,7 @@
 
 #include "cppbor.h"
 #include "cppbor_parse.h"
+#include "debug.hpp"
 #include "esp_log.h"
 #include "esp_timer.h"
 #include "ff.h"
@@ -66,8 +68,8 @@ static std::atomic<bool> sIsDbOpen(false);
 using std::placeholders::_1;
 using std::placeholders::_2;
 
-static auto CreateNewDatabase(leveldb::Options& options, locale::ICollator& col)
-    -> leveldb::DB* {
+static auto CreateNewDatabase(leveldb::Options& options,
+                              locale::ICollator& col) -> leveldb::DB* {
   Database::Destroy();
   leveldb::DB* db;
   options.create_if_missing = true;
@@ -348,6 +350,8 @@ auto Database::updateIndexes() -> void {
   }
   update_tracker_ = std::make_unique<UpdateTracker>();
 
+  tag_parser_.ClearCaches();
+
   leveldb::ReadOptions read_options;
   read_options.fill_cache = false;
   read_options.verify_checksums = true;
@@ -373,21 +377,24 @@ auto Database::updateIndexes() -> void {
         continue;
       }
 
+      std::shared_ptr<TrackTags> tags;
       FILINFO info;
       FRESULT res = f_stat(track->filepath.c_str(), &info);
-
-      std::pair<uint16_t, uint16_t> modified_at{0, 0};
       if (res == FR_OK) {
-        modified_at = {info.fdate, info.ftime};
-      }
-      if (modified_at == track->modified_at) {
-        continue;
-      } else {
-        track->modified_at = modified_at;
+        std::pair<uint16_t, uint16_t> modified_at{info.fdate, info.ftime};
+        if (modified_at == track->modified_at) {
+          continue;
+        } else {
+          // Will be written out later; we make sure we update the track's
+          // modification time and tags in the same batch so that interrupted
+          // reindexes don't cause a big mess.
+          track->modified_at = modified_at;
+        }
+
+        tags = tag_parser_.ReadAndParseTags(
+            {track->filepath.data(), track->filepath.size()});
       }
 
-      std::shared_ptr<TrackTags> tags = tag_parser_.ReadAndParseTags(
-          {track->filepath.data(), track->filepath.size()});
       if (!tags || tags->encoding() == Container::kUnsupported) {
         // We couldn't read the tags for this track. Either they were
         // malformed, or perhaps the file is missing. Either way, tombstone
@@ -411,30 +418,24 @@ auto Database::updateIndexes() -> void {
       // At this point, we know that the track still exists in its original
       // location. All that's left to do is update any metadata about it.
 
-      auto new_type = calculateMediaType(*tags, track->filepath);
-      uint64_t new_hash = tags->Hash();
-      if (new_hash != track->tags_hash || new_type != track->type) {
-        // This track's tags have changed. Since the filepath is exactly the
-        // same, we assume this is a legitimate correction. Update the
-        // database.
-        ESP_LOGI(kTag, "updating hash (%llx -> %llx)", track->tags_hash,
-                 new_hash);
-
-        // Again, we remove the old index records first so has to avoid
-        // dangling references.
-        dbRemoveIndexes(track);
+      // Remove the old index records first so has to avoid dangling records.
+      dbRemoveIndexes(track);
 
-        // Atomically correct the hash + create the new index records.
-        leveldb::WriteBatch batch;
-        track->tags_hash = new_hash;
-        dbIngestTagHashes(*tags, track->individual_tag_hashes, batch);
+      // Atomically correct the hash + create the new index records.
+      leveldb::WriteBatch batch;
 
-        track->type = new_type;
-        dbCreateIndexesForTrack(*track, *tags, batch);
-        batch.Put(EncodeDataKey(track->id), EncodeDataValue(*track));
+      uint64_t new_hash = tags->Hash();
+      if (track->tags_hash != new_hash) {
+        track->tags_hash = new_hash;
         batch.Put(EncodeHashKey(new_hash), EncodeHashValue(track->id));
-        db_->Write(leveldb::WriteOptions(), &batch);
       }
+
+      track->type = calculateMediaType(*tags, track->filepath);
+      batch.Put(EncodeDataKey(track->id), EncodeDataValue(*track));
+
+      dbIngestTagHashes(*tags, track->individual_tag_hashes, batch);
+      dbCreateIndexesForTrack(*track, *tags, batch);
+      db_->Write(leveldb::WriteOptions(), &batch);
     }
   }
 
@@ -445,8 +446,8 @@ auto Database::updateIndexes() -> void {
   track_finder_.launch("");
 };
 
-auto Database::processCandidateCallback(FILINFO& info, std::string_view path)
-    -> void {
+auto Database::processCandidateCallback(FILINFO& info,
+                                        std::string_view path) -> void {
   leveldb::ReadOptions read_options;
   read_options.fill_cache = true;
   read_options.verify_checksums = false;
@@ -530,9 +531,8 @@ static constexpr char kMusicMediaPath[] = "/Music/";
 static constexpr char kPodcastMediaPath[] = "/Podcasts/";
 static constexpr char kAudiobookMediaPath[] = "/Audiobooks/";
 
-auto Database::calculateMediaType(TrackTags& tags, std::string_view path)
-    -> MediaType {
-
+auto Database::calculateMediaType(TrackTags& tags,
+                                  std::string_view path) -> MediaType {
   auto equalsIgnoreCase = [&](char lhs, char rhs) {
     return std::tolower(lhs) == std::tolower(rhs);
   };
@@ -540,7 +540,8 @@ auto Database::calculateMediaType(TrackTags& tags, std::string_view path)
   // Use the filepath first, since it's the most explicit way for the user to
   // tell us what this track is.
   auto checkPathPrefix = [&](std::string_view path, std::string prefix) {
-    auto res = std::mismatch(prefix.begin(), prefix.end(), path.begin(), path.end(), equalsIgnoreCase);
+    auto res = std::mismatch(prefix.begin(), prefix.end(), path.begin(),
+                             path.end(), equalsIgnoreCase);
     return res.first == prefix.end();
   };
 
@@ -634,8 +635,8 @@ auto Database::dbMintNewTrackId() -> TrackId {
   return next_track_id_++;
 }
 
-auto Database::dbGetTrackData(leveldb::ReadOptions options, TrackId id)
-    -> std::shared_ptr<TrackData> {
+auto Database::dbGetTrackData(leveldb::ReadOptions options,
+                              TrackId id) -> std::shared_ptr<TrackData> {
   std::string key = EncodeDataKey(id);
   std::string raw_val;
   if (!db_->Get(options, key, &raw_val).ok()) {
@@ -668,26 +669,55 @@ auto Database::dbRemoveIndexes(std::shared_ptr<TrackData> data) -> void {
   }
   for (const IndexInfo& index : getIndexes()) {
     auto entries = Index(collator_, index, *data, *tags);
+    std::optional<uint8_t> preserve_depth{};
+
+    // Iterate through the index records backwards, so that we start deleting
+    // records from the highest depth. This allows us to work out what depth to
+    // stop at as we go.
+    // e.g. if the last track of an album is being deleted, we need to delete
+    // the album index record at n-1 depth. But if there are still other tracks
+    // in the album, then we should only clear records at depth n.
     for (auto it = entries.rbegin(); it != entries.rend(); it++) {
+      if (preserve_depth) {
+        if (it->first.header.components_hash.size() < *preserve_depth) {
+          // Some records deeper than this were not removed, so don't try to
+          // remove this one.
+          continue;
+        }
+        // A record weren't removed, but the current record is either a
+        // sibling of that record, or a leaf belonging to a sibling of that
+        // record. Either way, we can start deleting records again.
+        // preserve_depth will be set to a new value if we start encountering
+        // siblings again.
+        preserve_depth.reset();
+      }
+
       auto key = EncodeIndexKey(it->first);
       auto status = db_->Delete(leveldb::WriteOptions{}, key);
       if (!status.ok()) {
         return;
       }
 
+      // Are any sibling records left at this depth? If two index records have
+      // matching headers, they are siblings (same depth, same selections at
+      // lower depths)
       std::unique_ptr<leveldb::Iterator> cursor{db_->NewIterator({})};
+
+      // Check the record before the one we just deleted.
       cursor->Seek(key);
       cursor->Prev();
-
       auto prev_key = ParseIndexKey(cursor->key());
       if (prev_key && prev_key->header == it->first.header) {
-        break;
+        preserve_depth = it->first.header.components_hash.size();
+        continue;
       }
 
+      // Check the record after.
       cursor->Next();
       auto next_key = ParseIndexKey(cursor->key());
       if (next_key && next_key->header == it->first.header) {
-        break;
+        preserve_depth = it->first.header.components_hash.size();
+        continue;
       }
     }
   }
@@ -809,8 +839,7 @@ Iterator::Iterator(std::shared_ptr<Database> db, IndexId idx)
     : Iterator(db,
                IndexKey::Header{
                    .id = idx,
-                   .depth = 0,
-                   .components_hash = 0,
+                   .components_hash = {},
                }) {}
 
 Iterator::Iterator(std::shared_ptr<Database> db, const IndexKey::Header& header)
@@ -883,8 +912,8 @@ auto TrackIterator::next() -> void {
 
     auto& cur = levels_.back().value();
     if (!cur) {
-      // The current top iterator is out of tracks. Pop it, and move the parent
-      // to the next item.
+      // The current top iterator is out of tracks. Pop it, and move the
+      // parent to the next item.
       levels_.pop_back();
     } else if (std::holds_alternative<IndexKey::Header>(cur->contents())) {
       // This record is a branch. Push a new iterator.
diff --git a/src/tangara/database/database.hpp b/src/tangara/database/database.hpp
index ef96f859..6d91b04c 100644
--- a/src/tangara/database/database.hpp
+++ b/src/tangara/database/database.hpp
@@ -38,7 +38,7 @@
 
 namespace database {
 
-const uint8_t kCurrentDbVersion = 9;
+const uint8_t kCurrentDbVersion = 10;
 
 struct SearchKey;
 class Record;
diff --git a/src/tangara/database/index.cpp b/src/tangara/database/index.cpp
index e60f56d5..7f594700 100644
--- a/src/tangara/database/index.cpp
+++ b/src/tangara/database/index.cpp
@@ -77,8 +77,8 @@ const IndexInfo kAudiobooks{
     .components = {Tag::kAlbum, Tag::kAlbumOrder},
 };
 
-static auto titleOrFilename(const TrackData& data, const TrackTags& tags)
-    -> std::pmr::string {
+static auto titleOrFilename(const TrackData& data,
+                            const TrackTags& tags) -> std::pmr::string {
   auto title = tags.title();
   if (title) {
     return *title;
@@ -147,8 +147,7 @@ auto Indexer::index() -> std::vector<std::pair<IndexKey, std::string>> {
 
   IndexKey::Header root_header{
       .id = index_.id,
-      .depth = 0,
-      .components_hash = 0,
+      .components_hash = {},
   };
   handleLevel(root_header, index_.components);
 
@@ -237,16 +236,10 @@ auto Index(locale::ICollator& collator,
 }
 
 auto ExpandHeader(const IndexKey::Header& header,
-                  const std::optional<std::pmr::string>& component)
-    -> IndexKey::Header {
+                  const std::pmr::string& component) -> IndexKey::Header {
   IndexKey::Header ret{header};
-  ret.depth++;
-  if (component) {
-    ret.components_hash =
-        komihash(component->data(), component->size(), ret.components_hash);
-  } else {
-    ret.components_hash = komihash(NULL, 0, ret.components_hash);
-  }
+  ret.components_hash.push_back(
+      komihash(component.data(), component.size(), 0));
   return ret;
 }
 
diff --git a/src/tangara/database/index.hpp b/src/tangara/database/index.hpp
index 98e57d54..c4dd5701 100644
--- a/src/tangara/database/index.hpp
+++ b/src/tangara/database/index.hpp
@@ -41,14 +41,8 @@ struct IndexKey {
   struct Header {
     // The index that this key was created for.
     IndexId id;
-    // The number of components of IndexInfo that have already been filtered.
-    // For example, if an index consists of { kGenre, kArtist }, and this key
-    // represents an artist, then depth = 1.
-    std::uint8_t depth;
-    // The cumulative hash of all filtered components, in order. For example, if
-    // an index consists of { kArtist, kAlbum, kTitle }, and we are at depth = 2
-    // then this may contain hash(hash("Jacqueline"), "My Cool Album").
-    std::uint64_t components_hash;
+    // The hashes of all filtered components, in order, up to the current depth
+    std::vector<std::uint64_t> components_hash;
 
     bool operator==(const Header&) const = default;
   };
@@ -56,7 +50,7 @@ struct IndexKey {
 
   // The filterable / selectable item that this key represents. "Jacqueline" for
   // kArtist, "My Cool Album" for kAlbum, etc.
-  std::optional<std::pmr::string> item;
+  std::pmr::string item;
   // If this is a leaf component, the track id for this record.
   // This could reasonably be the value for a record, but we keep it as a part
   // of the key to help with disambiguation.
@@ -69,7 +63,7 @@ auto Index(locale::ICollator&,
            const TrackTags&) -> std::vector<std::pair<IndexKey, std::string>>;
 
 auto ExpandHeader(const IndexKey::Header&,
-                  const std::optional<std::pmr::string>&) -> IndexKey::Header;
+                  const std::pmr::string&) -> IndexKey::Header;
 
 // Predefined indexes
 // TODO(jacqueline): Make these defined at runtime! :)
diff --git a/src/tangara/database/records.cpp b/src/tangara/database/records.cpp
index addcc13d..852eb131 100644
--- a/src/tangara/database/records.cpp
+++ b/src/tangara/database/records.cpp
@@ -181,11 +181,13 @@ auto EncodeAllIndexesPrefix() -> std::string {
 auto EncodeIndexPrefix(const IndexKey::Header& header) -> std::string {
   std::ostringstream out;
   out << makePrefix(kIndexPrefix);
-  cppbor::Array val{
-      cppbor::Uint{header.id},
-      cppbor::Uint{header.depth},
-      cppbor::Uint{header.components_hash},
-  };
+
+  cppbor::Array components{};
+  for (auto hash : header.components_hash) {
+    components.add(cppbor::Uint{hash});
+  }
+
+  cppbor::Array val{cppbor::Uint{header.id}, std::move(components)};
   out << val.toString() << kFieldSeparator;
   return out.str();
 }
@@ -210,8 +212,8 @@ auto EncodeIndexKey(const IndexKey& key) -> std::string {
   out << EncodeIndexPrefix(key.header);
 
   // The component should already be UTF-8 encoded, so just write it.
-  if (key.item) {
-    out << *key.item << kFieldSeparator;
+  if (!key.item.empty()) {
+    out << key.item << kFieldSeparator;
   }
 
   if (key.track) {
@@ -236,14 +238,16 @@ auto ParseIndexKey(const leveldb::Slice& slice) -> std::optional<IndexKey> {
     return {};
   }
   auto as_array = key->asArray();
-  if (as_array->size() != 3 || as_array->get(0)->type() != cppbor::UINT ||
-      as_array->get(1)->type() != cppbor::UINT ||
-      as_array->get(2)->type() != cppbor::UINT) {
+  if (as_array->size() != 2 || as_array->get(0)->type() != cppbor::UINT ||
+      as_array->get(1)->type() != cppbor::ARRAY) {
     return {};
   }
   result.header.id = as_array->get(0)->asUint()->unsignedValue();
-  result.header.depth = as_array->get(1)->asUint()->unsignedValue();
-  result.header.components_hash = as_array->get(2)->asUint()->unsignedValue();
+  auto components_array = as_array->get(1)->asArray();
+  for (int i = 0; i < components_array->size(); i++) {
+    result.header.components_hash.push_back(
+        components_array->get(i)->asUint()->unsignedValue());
+  }
 
   size_t header_length =
       reinterpret_cast<const char*>(end_of_key) - key_data.data();
diff --git a/src/tangara/database/tag_parser.cpp b/src/tangara/database/tag_parser.cpp
index 6cdf6175..59bd8f13 100644
--- a/src/tangara/database/tag_parser.cpp
+++ b/src/tangara/database/tag_parser.cpp
@@ -190,6 +190,12 @@ auto TagParserImpl::ReadAndParseTags(std::string_view path)
     }
   }
 
+  // There wasn't a fine-grained 'Artists' tag in the source. Synthesize a tag
+  // based on the singlet 'Artist' field.
+  if (tags->allArtists().empty() && tags->artist()) {
+    tags->singleAllArtists(tags->artist().value());
+  }
+
   // Store the result in the cache for later.
   {
     std::lock_guard<std::mutex> lock{cache_mutex_};
@@ -199,6 +205,11 @@ auto TagParserImpl::ReadAndParseTags(std::string_view path)
   return tags;
 }
 
+auto TagParserImpl::ClearCaches() -> void {
+  std::lock_guard<std::mutex> lock{cache_mutex_};
+  cache_.Clear();
+}
+
 OggTagParser::OggTagParser() {}
 
 auto OggTagParser::ReadAndParseTags(std::string_view p)
diff --git a/src/tangara/database/tag_parser.hpp b/src/tangara/database/tag_parser.hpp
index 69b71940..220339c0 100644
--- a/src/tangara/database/tag_parser.hpp
+++ b/src/tangara/database/tag_parser.hpp
@@ -19,6 +19,8 @@ class ITagParser {
   virtual ~ITagParser() {}
   virtual auto ReadAndParseTags(std::string_view path)
       -> std::shared_ptr<TrackTags> = 0;
+
+  virtual auto ClearCaches() -> void {}
 };
 
 class TagParserImpl : public ITagParser {
@@ -27,6 +29,8 @@ class TagParserImpl : public ITagParser {
   auto ReadAndParseTags(std::string_view path)
       -> std::shared_ptr<TrackTags> override;
 
+  auto ClearCaches() -> void override;
+
  private:
   std::vector<std::unique_ptr<ITagParser>> parsers_;
 
@@ -58,14 +62,8 @@ class GenericTagParser : public ITagParser {
   // Supported file extensions for parsing tags, derived from the list of
   // supported audio formats here:
   // https://cooltech.zone/tangara/docs/music-library/
-  static constexpr std::string supported_exts[] = {
-    "flac",
-    "mp3",
-    "ogg",
-    "ogx",
-    "opus",
-    "wav"
-  };
+  static constexpr std::string supported_exts[] = {"flac", "mp3",  "ogg",
+                                                   "ogx",  "opus", "wav"};
 };
 
 }  // namespace database
diff --git a/src/tangara/database/track.cpp b/src/tangara/database/track.cpp
index ad9db1ba..dc66c9c9 100644
--- a/src/tangara/database/track.cpp
+++ b/src/tangara/database/track.cpp
@@ -83,11 +83,17 @@ auto tagToString(const TagValue& val) -> std::string {
           return std::to_string(arg);
         } else if constexpr (std::is_same_v<
                                  T, std::span<const std::pmr::string>>) {
-          std::ostringstream builder{};
+          std::stringstream stream{};
+          bool first = true;
           for (const auto& str : arg) {
-            builder << std::string{str.data(), str.size()} << ",";
+            if (first) {
+              first = false;
+            } else {
+              stream << ";";
+            }
+            stream << std::string{str.data(), str.size()};
           }
-          return builder.str();
+          return stream.str();
         }
       },
       val);
@@ -101,7 +107,6 @@ auto tagToString(const TagValue& val) -> std::string {
 auto parseDelimitedTags(const std::string_view s,
                         const char* delimiters,
                         std::pmr::vector<std::pmr::string>& out) -> void {
-  out.clear();
   std::string src = {s.data(), s.size()};
   char* token = std::strtok(src.data(), delimiters);
 
@@ -240,7 +245,6 @@ auto TrackTags::artist() const -> const std::optional<std::pmr::string>& {
 
 auto TrackTags::artist(std::string_view s) -> void {
   artist_ = s;
-  maybeSynthesizeAllArtists();
 }
 
 auto TrackTags::allArtists() const -> std::span<const std::pmr::string> {
@@ -249,7 +253,10 @@ auto TrackTags::allArtists() const -> std::span<const std::pmr::string> {
 
 auto TrackTags::allArtists(const std::string_view s) -> void {
   parseDelimitedTags(s, kAllArtistDelimiters, allArtists_);
-  maybeSynthesizeAllArtists();
+}
+
+auto TrackTags::singleAllArtists(const std::string_view s) -> void {
+  allArtists_.emplace_back(s);
 }
 
 auto TrackTags::album() const -> const std::optional<std::pmr::string>& {
@@ -325,17 +332,6 @@ auto TrackTags::Hash() const -> uint64_t {
   return komihash_stream_final(&stream);
 }
 
-/*
- * Adds the current 'artist' tag to 'allArtists' if needed. Many tracks lack a
- * fine-grained 'ARTISTS=' tag (or equivalent), but pushing down this nuance to
- * consumers of TrackTags adds a lot of complexity.
- */
-auto TrackTags::maybeSynthesizeAllArtists() -> void {
-  if (allArtists_.empty() && artist_) {
-    allArtists_.push_back(*artist_);
-  }
-}
-
 auto database::TrackData::clone() const -> std::shared_ptr<TrackData> {
   auto data = std::make_shared<TrackData>();
   data->id = id;
diff --git a/src/tangara/database/track.hpp b/src/tangara/database/track.hpp
index 71eb44ce..c7dff425 100644
--- a/src/tangara/database/track.hpp
+++ b/src/tangara/database/track.hpp
@@ -117,6 +117,7 @@ class TrackTags {
 
   auto allArtists() const -> std::span<const std::pmr::string>;
   auto allArtists(const std::string_view) -> void;
+  auto singleAllArtists(const std::string_view) -> void;
 
   auto album() const -> const std::optional<std::pmr::string>&;
   auto album(std::string_view) -> void;
@@ -144,8 +145,6 @@ class TrackTags {
   auto Hash() const -> uint64_t;
 
  private:
-  auto maybeSynthesizeAllArtists() -> void;
-
   Container encoding_;
 
   std::optional<std::pmr::string> title_;
diff --git a/src/tangara/lua/lua_database.cpp b/src/tangara/lua/lua_database.cpp
index e79d6141..928dbc39 100644
--- a/src/tangara/lua/lua_database.cpp
+++ b/src/tangara/lua/lua_database.cpp
@@ -190,35 +190,24 @@ static const struct luaL_Reg kDatabaseFuncs[] = {
     {"update", update},   {"track_by_id", track_by_id},
     {NULL, NULL}};
 
-/*
- * Struct to be used as userdata for the Lua representation of database records.
- * In order to push these large values into PSRAM as much as possible, memory
- * for these is allocated and managed by Lua. This struct must therefore be
- * trivially copyable.
- */
-struct LuaRecord {
-  std::variant<database::TrackId, database::IndexKey::Header> contents;
-  size_t text_size;
-  char text[];
-};
-
-static_assert(std::is_trivially_destructible<LuaRecord>());
-static_assert(std::is_trivially_copy_assignable<LuaRecord>());
-
-static auto push_lua_record(lua_State* L, const database::Record& r) -> void {
-  // Create and init the userdata.
-  LuaRecord* record = reinterpret_cast<LuaRecord*>(
-      lua_newuserdata(L, sizeof(LuaRecord) + r.text().size()));
-  luaL_setmetatable(L, kDbRecordMetatable);
+static auto push_lua_record(lua_State* state,
+                            const database::Record& r) -> void {
+  database::Record** data = reinterpret_cast<database::Record**>(
+      lua_newuserdata(state, sizeof(uintptr_t)));
+  *data = new database::Record(r);
+  luaL_setmetatable(state, kDbRecordMetatable);
+}
 
-  // Init all the fields
-  *record = {
-      .contents = r.contents(),
-      .text_size = r.text().size(),
-  };
+auto db_check_record(lua_State* L, int stack_pos) -> database::Record* {
+  database::Record* rec = *reinterpret_cast<database::Record**>(
+      luaL_checkudata(L, stack_pos, kDbRecordMetatable));
+  return rec;
+}
 
-  // Copy the string data across.
-  std::memcpy(record->text, r.text().data(), r.text().size());
+static auto db_record_gc(lua_State* state) -> int {
+  database::Record* rec = db_check_record(state, 1);
+  delete rec;
+  return 0;
 }
 
 auto db_check_iterator(lua_State* L, int stack_pos) -> database::Iterator* {
@@ -227,8 +216,8 @@ auto db_check_iterator(lua_State* L, int stack_pos) -> database::Iterator* {
   return it;
 }
 
-static auto push_iterator(lua_State* state, const database::Iterator& it)
-    -> void {
+static auto push_iterator(lua_State* state,
+                          const database::Iterator& it) -> void {
   database::Iterator** data = reinterpret_cast<database::Iterator**>(
       lua_newuserdata(state, sizeof(uintptr_t)));
   *data = new database::Iterator(it);
@@ -279,15 +268,13 @@ static const struct luaL_Reg kDbIteratorFuncs[] = {
     {"__gc", db_iterator_gc},     {NULL, NULL}};
 
 static auto record_text(lua_State* state) -> int {
-  LuaRecord* data = reinterpret_cast<LuaRecord*>(
-      luaL_checkudata(state, 1, kDbRecordMetatable));
-  lua_pushlstring(state, data->text, data->text_size);
+  database::Record* rec = db_check_record(state, 1);
+  lua_pushlstring(state, rec->text().data(), rec->text().size());
   return 1;
 }
 
 static auto record_contents(lua_State* state) -> int {
-  LuaRecord* data = reinterpret_cast<LuaRecord*>(
-      luaL_checkudata(state, 1, kDbRecordMetatable));
+  database::Record* rec = db_check_record(state, 1);
 
   std::visit(
       [&](auto&& arg) {
@@ -304,7 +291,7 @@ static auto record_contents(lua_State* state) -> int {
           }
         }
       },
-      data->contents);
+      rec->contents());
 
   return 1;
 }
@@ -312,6 +299,7 @@ static auto record_contents(lua_State* state) -> int {
 static const struct luaL_Reg kDbRecordFuncs[] = {{"title", record_text},
                                                  {"contents", record_contents},
                                                  {"__tostring", record_text},
+                                                 {"__gc", db_record_gc},
                                                  {NULL, NULL}};
 
 static auto index_name(lua_State* state) -> int {
@@ -405,7 +393,6 @@ static auto lua_database(lua_State* state) -> int {
   lua_rawset(state, -3);
   lua_rawset(state, -3);
 
-
   lua_pushliteral(state, "IndexTypes");
   lua_newtable(state);
   lua_pushliteral(state, "ALBUMS_BY_ARTIST");
diff --git a/src/tangara/lua/lua_database.hpp b/src/tangara/lua/lua_database.hpp
index 51e71758..fa2c26a3 100644
--- a/src/tangara/lua/lua_database.hpp
+++ b/src/tangara/lua/lua_database.hpp
@@ -13,8 +13,9 @@
 namespace lua {
 
 auto db_check_iterator(lua_State*, int stack_pos) -> database::Iterator*;
+auto db_check_record(lua_State*, int stack_pos) -> database::Record*;
 
-auto pushTagValue(lua_State* L, const database::TagValue& val) -> void; 
+auto pushTagValue(lua_State* L, const database::TagValue& val) -> void;
 
 auto RegisterDatabaseModule(lua_State*) -> void;