Skip to content

counterparts 0.3.27

Choose a tag to compare

@mlapeter mlapeter released this 19 Oct 19:42
· 1735 commits to master since this release

0.3.27 — 2026-10-19

Recall learns from use, and finds "baking" when you say "bake". The memories a session rates useful, draws on or opens count as used, memories found useful together get linked, and one rated noise twice comes up less for a while; what the conversation has been about, and what is open in its folder, gives related memories a head start. Recall matches word forms and counts everyday words for less, a session's first prompt is matched by meaning too, footnotes skip what this session wrote, and a newer memory stands in for the one it replaced. Most of these are experiments, to be checked in a few days. counterparts doctor gains a Feedback loop line. No store migration (store v13) and no wiring change, so counterparts connect is not needed; the word index re-indexes itself once in the background (cache v6 to v7, under a second on a typical store). Going back to 0.3.26 is not a plain reinstall: its memory tools refuse the newer cache, and its hooks miss word matches without saying so. To go back, delete ~/.counterparts/store/cache/cache.sqlite*, then run counterparts verify --rebuild --keep-vectors --dir ~/.counterparts/store on 0.3.26 (recall by meaning comes back over the next few sessions, as the background worker re-reads the memories); counterparts rollback does not cover a cache change.

  • Recall by meaning works on a session's first prompt. Until now the first prompt of a
    session, and any prompt whose background cue from the turn before wasn't ready, was
    matched by words only. Those prompts are now matched by meaning too, computed on the spot
    with the same local table. It adds a few tens of milliseconds on a typical store.
  • Asking for facts ignores "how many" and "how often" wording. Words like "many",
    "much", "often", "typical" and "usually" no longer count as part of what the question is
    about, so they stop pulling in memories that merely contain them.
  • Footnotes no longer repeat what this session just wrote (an experiment). A memory
    written in the current conversation is already in front of the model, so it no longer
    comes back as a footnote or a "came to mind" line there. After the conversation is
    compacted, what it wrote before the compaction can come back again. Asking for a memory on
    purpose still finds it.
  • The nightly dream replays what the day used (an experiment). Memories the day opened,
    quoted, drew on or rated useful are handed to the night's dream, even ones it already
    dreamed over, with the ones that changed what was said first, and pairs used together
    listed for it to link.
  • The nightly run's own sessions get no footnotes (an experiment). The background
    sessions that dream and reflect overnight are no longer handed memories at their first
    prompt; they read what they need on purpose.
  • Rating what you were shown now counts, as an experiment. When a session rates the
    memories it was shown at the end, a 3 counts as a full use of the memory, a 2 most of one,
    a 1 a little, and a footnote left unrated a glimpse; a 0 counts nothing. Used memories fade
    more slowly, as before. This is deliberately strong so its effect can be seen, and will be
    tuned after a few days.
  • A memory a session says it drew on counts as a full use, as an experiment. Naming it
    in drewOn now counts like opening it (it was half a use); still once a day, at most three
    per boundary.
  • Memories found useful together get linked strongly. Two memories rated useful (or
    opened, quoted or drawn on) since the session last rated are linked strongly enough that
    the second time it happens, one can bring the other to mind as a quiet pointer. Memories
    shown on the same turn get a faint link, unless one was rated noise.
  • A memory rated noise twice comes up less for a while. Rated 0 in two sessions within
    a week, it needs a closer match to show as a footnote. It recovers when those ratings are
    a week old, or at once when a later session rates it useful.
  • counterparts doctor has a Feedback loop line with what the ratings credited by
    score, the pairs they linked, the links now strong enough to show as pointers, and how
    many memories are holding back as noise.
  • An earlier version of a memory no longer shows without the one that replaced it. When
    a turn's words lean towards a memory that a newer one has settled over, the newer one now
    shows in its place (labelled with the earlier one), where before the earlier one could
    slip through on its own as a quiet footnote. If the newer one doesn't clear the bar, the
    earlier one stays out too. If the newer one was written in this same
    session, neither is shown, since it's already in the conversation. This is a trial, to be
    checked in a few days.
  • A new memory is no longer linked to what was shown when it was written, as an
    experiment.
    These faint links were almost never confirmed by use and mostly noise, and
    memories found useful together are now linked by ratings instead. Links between memories
    written one after another stay, and the links already made fade on their own.
  • A memory that looks like an older one's update no longer fades the older one right away,
    as an experiment.
    It is still marked "may be out of date" and the nightly dream is asked
    to confirm it; only a confirmation fades it. The guess was wrong too often to cost the
    older memory anything before then.
  • Recall matches word forms and quiets everyday words (an experiment). A prompt that
    says "bake" now finds a memory that says "baking" or "baked", and words like "any",
    "what" or "you" count for much less when deciding what a prompt is about. This applies to
    recall as you work and to asking for facts. The first time the background worker runs
    after the update it re-indexes the words of your memories (under a second on a typical
    store); until then recall matches both the old and the new word forms.
  • Recall follows the conversation's train of thought (an experiment). What a
    conversation has been about, what it opened, rated useful or wrote, and what was left
    open in its folder now give related memories a head start, so they win close calls they
    used to lose. Only memories the current message already touches can get one, a memory
    the conversation wrote itself is still never handed back, and the head start fades
    within a few messages once the topic moves on. A conversation's first message, with
    nothing open in its folder, is recalled exactly as before.