v4.100.265

Memory and the Library

• A message with the word "memory" no longer recalls every note. Each note
  was indexed with its internal name, memory:<hash>, as if it were a file
  name, so every note held the word "memory" and any search with it matched
  them all. Notes kept before this version are indexed again the first time
  Memory opens.
• With a reranker set, automatic recall attaches only notes it scores high.
  The reranker's score passed every note, because the relevance threshold
  defaults to 0, so the five best notes went with every message. A reranked
  note is now attached when it scores 0.5 or more, or when it is close by
  meaning or shares two meaningful words with the message, as without a
  reranker.
• A recall that cannot run says so. When Memory cannot be searched, the
  task shows "Memory unreachable, the message went without notes: ",
  once per task for each reason. Until now it read the same as Memory having
  nothing to say. When the search ran by keyword alone because LanceDB is not
  installed, recall and search_library say that too.
• Recalled notes take at most 4% of the model's window. Five notes of up to
  4000 characters each could take a fifth of a 32k window. The best note is
  always given, cut to fit if it has to be, with its id to recall it whole;
  the task notice says how many more did not fit.
• A note a compaction folded away is recalled again when it comes up. A
  task was given each note once, and a note summarised away by a compaction
  never came back that task. Notes still in the conversation word for word are
  not given twice.
• The same note with other spacing or capitals is kept once, and noting it
  again with new tags adds the tags. Notes that differ only that way are
  merged when Memory opens: the oldest keeps its id, with the tags of all of
  them.
• A note's id no longer changes. A note noted again under other chunk
  settings was replaced under a new id, and m<id> references to it stopped
  working.
• A re-pulled embedding model is noticed, and every collection catches up.
  Adding a book embedded only that book; the rest stayed without vectors for a
  new model until something was added to each. Now the rest follows in the
  background. The model is also checked by the vector it gives for a fixed
  text, so a model re-pulled under the same name and size, whose vectors no
  longer compare with the stored ones, has its old vectors dropped and
  everything embedded again.
• Two embedding runs at once, such as a book added while an import is still
  embedding, no longer write a document's vectors twice or fail on a LanceDB
  commit conflict.
• Notes say what they are and where they stand. A note is a fact, a
  decision (with the reason it was taken) or a preference, and it can name the
  places in the code it is about, as path or path:line. The model keeps
  notes as noted. Only you can confirm one or mark it outdated, and either
  takes a reason, which the model is shown with the note. Each recalled note carries a line telling the model what to do with
  it. A confirmed decision is not reopened unless something contradicts its
  reason. A noted fact is checked before it is relied on. An outdated note is
  not attached again.
• A notes browser in Settings → Memory. The Notes button beside each
  memory lists its notes by status, with evidence and earlier wording. From
  there you can Confirm… a note, mark it Outdated…, put it Back to noted or
  Forget it, and Merge near-duplicates into the oldest note. The merge also
  runs on its own, once an hour while Memory is on. A confirmed note is
  never merged into one that is not.
• A changed note is revised in place. It keeps its id, the earlier wording
  is kept with it, and its status goes back to noted. A note id is never given
  to another note, even after the note is forgotten.
• A note that says nearly the same as one already kept is refused, naming
  the note it repeats, so the model revises that one instead of keeping a
  second copy.
• Recall ranks by use as well as by match. A note given lately counts a
  little more. One given at every message, whatever was asked, counts less.
• Notes about the files a task is working on come up with it. Up to three
  notes whose places match the files the task has read or edited are
  attached, even when the message does not mention them. They are matched by
  place, never by words alone.
• A note's places are checked when it is given. When the line it named
  has moved, changed or gone, or the path matches more than one file, the
  note says so.
• Facts can expire. A new setting, "Facts expire after (days)", is off by
  default. Past it, a noted fact is distilled with the others that expired
  into one note tagged distilled, and the originals are removed only after
  that note is written. Confirmed notes, decisions and preferences never
  expire. The distilling model is the profile picked for Memory.

Compaction

• A compaction summary no longer narrates steps that were never taken.
  A writer whose part of the conversation ended mid-task could carry on in the
  present tense, citing calls past the end of the record. claude-hooks measured
  citations up to cline#174 against 55 calls with qwen3.6 27B. Such an attempt is
  now asked again with the numbers it invented. What is still left after the
  last attempt, or that a reviewer brought back, is cut clause by clause:
  never past the previous citation and never across a paragraph. Until now
  only the bad number was dropped, and the invented sentence stayed.
• The writer reads each tool call with its number. The conversation it
  summarises now shows #7 read_files(...), the same number the call has in
  the record, so the writer cites a call by reading its number instead of
  counting calls down the page.
• A summary no longer states details of files nothing opened. On a first
  compaction, writers and reviewers filled in the task's module tables for
  files that had only been listed, with every field invented (1 right in 98,
  in claude-hooks' measurement). A summary line whose every file is one that
  no call showed is now cut. The files are named in one "Not yet read" note
  at the end, so a list of what is still to read survives as names. A
  listing does not count as having read a file. A file the user's own
  requests name is never cut, and fenced code is left alone. The set of files
  seen is carried from one compaction to the next.

Tools

• check_file names the end of a file in the file's own count. For an
  unclosed bracket the language server reports where the parser gave up,
  which is the end of the file. It numbers that as the empty line after the
  last line break, a line no other tool counts. A page that read_files
  called 136 lines long got Line 137, column 1: '}' expected. The model
  took the line that did not exist for a check of an old copy, and stopped
  trusting the check. Such a position now reads as "the end of the file
  (after line 136, its last…)". For a page it adds the line of the script's
  </script>.
• A column edit whose old_text is not at those columns is no longer
  applied. The editor replaced the columns it was given and ignored
  old_text. Asked to delete a } at a column that held the l of else,
  it deleted the l and reported success. Now, when old_text occurs once
  on the lines named, the edit goes there and the result says so. Otherwise
  it is refused, naming what the columns hold and every column where
  old_text does occur.
• The check_file description in the SDK and in every prompt template now
  matches the plugin's. It was missing the sentence on finding a bracket by
  the text before it rather than by its column.
• An edit that sends its own old_text back as new_text is told so.
  A small model retyping a long minified line copies it unchanged, with the
  change it meant left out. The refusal used to say that the call was not
  landing where it thought and to try columns, though it was landing exactly
  there. It now says the change was never written into new_text. On a long
  line it shows how to send only the characters that change, with the line
  they are on.
• An edit card shows which characters really change. On the added
  lines of an edit card, characters the new text shares with the old are
  yellow, and only what is actually new stays green. Removed lines stay all
  red. An edit that sends the old text back unchanged shows its added line
  solid yellow, not green.
• An edit card says whether the edit was applied. A refused edit came
  back as a result carrying an error, not as a tool failure, so it drew the
  same "wants to edit this file" card as one that landed. Once the editor
  has answered, the card now reads "edited this file" or "edit was refused",
  in red when refused. It carries the editor's answer: Result, or Refused
  with the editor's own words, such as where old_text diverged from the
  file. It shows its first line and expands to the whole answer. The same
  goes for the cards of new and deleted files, and for tasks reopened from
  history.
• A landed edit clears the repeat guard's verdicts on that file. A call
  refused as a no-op, or failed, used to stay on its countdown even after
  other edits changed the same file, so a call that had become right could
  still be warned or stopped. Now an edit that lands on a file clears those
  verdicts for every call on that file. A call that keeps getting the same
  answer is still stopped as before, which is what catches a model
  restoring the file and reapplying the same edit over and over.
• An edit right after restore_file is no longer refused as unread.
  A restore erased the record that the file had been read, so the next
  edit, even one matched on text, was refused with "has not been read in
  this session". Now a restore only marks the old line numbers stale. An
  edit by line still asks for a fresh read, and says why. An edit matched
  on text goes through. The same applies after a transaction rollback and
  after an expert's edits.
• Running a command again is never flagged as a repeat. The repeat guard
  warned when a model ran the same command twice. It flagged a task's own
  check, such as a game launcher re-run after each edit, and told the model
  to change approach when it was doing what it should. A command's answer
  depends on the files, not on its arguments, so run_commands is no longer
  counted at all. It still breaks another call's run of repeats.
• The delimiter scan suggests the missing bracket, not a move. When a
  line was one } short and a spare } sat further down the file, the scan
  suggested moving a ) instead of adding the }. That only balanced by
  using up the spare, and left the code still broken. Now a fix that
  balances the line's own count comes first, and the spare is reported
  separately.