Skip to content

Releases

counterparts 0.3.54

Choose a tag to compare

@mlapeter mlapeter released this 10 Nov 18:46

Hosted gets your dashboard (read-only: home, memories, your page and dreams), and hosted sign-up can open by spots instead of invite codes (closed until we open it). What you'll notice on hosted: the account page links "Open your dashboard", which shows Home, Memories and Self; Subjects, Flow and Health say they're not on hosted yet. The sign-in page reads "Email me a link", with an optional "Have an invite code?". For local users: the dashboard's buttons work as before (their click handling moved, nothing new). No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.53 is a reinstall with no store restore: counterparts disconnect, 0.3.53's install, then counterparts connect.

  • Hosted: your dashboard, read-only. The account page links "Open your dashboard" (/dashboard/): the local
    dashboard's Home, Memories and Self pages (the days, the memories, the self page, the dreams), served from your
    hosted memory and seen only by you. Nothing on it writes; the pages that read your computer (Subjects, Flow,
    Health) say they're not on hosted yet. The local dashboard and hosted's now share one server-side answerer, and
    every button's click handling moved from inline handlers to delegated listeners: the local dashboard looks and
    works as before. Its pages retry a request that was told to slow down.
  • Hosted: sign up with just your email while the beta has spots left. The operator sets how many accounts
    hosted takes (spots set N); until it's full, a new address gets a sign-up link with no code. When it's full,
    the sign-in page says so and links counterparts.ai's waiting list. Invite codes still work, and are now an
    optional "Have an invite code?" field. Spots start at 0, so nothing opens until the operator says so.

counterparts 0.3.53

Choose a tag to compare

@mlapeter mlapeter released this 10 Nov 14:30

Your page names what it sets down and tells more than corrections; the morning wake reads cleaner; things waiting on you come up when they're relevant, not as a pile. What you'll notice: the morning after a night's page write, the wake can say "Set down from my page last night: …"; the wake ends with a "Carried forward" line, gives overdue items as a count with the oldest date, and names the session id in its first line; and now and then a line "Waiting on , ask when it fits: …" when something you owe an answer on comes up in the conversation or at the week's session start. The wake changes are in the shaped wake (on for a store created since 0.3.46, or after counterparts wake-shape on). remember takes story: true. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.52 is a reinstall with no store restore: counterparts disconnect, 0.3.52's install, then counterparts connect.

  • The page says what it sets down. The self page stays within its room, so each night a new line pushes an old
    one out. Now the night's page writers name the lines they take out, with a short why, and anything taken out
    that wasn't named is recorded anyway; a line only reworded, or folded into a story still on the page, isn't
    counted as gone. Each line set down is kept as a memory, word for word. The next morning's wake names up to two
    ("Set down from my page last night: …"), and recall "set down" lists them all, newest first, with why and date.
    When life brings one up again (it's used, or a newer memory says it again), the night's writers are shown it and
    may put it back.
  • The page's stories are about growth, not only corrections. On a store that keeps its page in its own form
    (counterparts page-form own), the night's page writers are told to keep a mix of story kinds (corrections that
    changed how the counterpart works, and also what went right, discoveries, and moments with the owner), and that a
    story earns its place if it would still teach something in a different project. Day's end is shown the day's
    memories about me or us with a small head start over the day's work, so a busy work day doesn't crowd them out.
    A practice on the page that is only about the work may move into a skill memory, which recall brings back when
    the work calls for it; the set-down record names it. On every store, remember takes story: true to nominate a
    moment as a story: the night's writer sees it first and still decides, and one it passes by fades like any
    other memory.
  • The wake says what it points at, once. A reminder told in another session is listed again by title instead of
    "shown earlier", which that session never saw. The night's closing line no longer repeats a sentence the page or
    day's end already says, and the same line on a second morning shrinks to a clause. Overdue items get their own count
    with the oldest date. A handoff waiting on an answer with nothing new sinks below the others, and one more than a week
    old that nothing refers to stops being listed (still under recall "handoffs"). Every line under Elsewhere says when.
    "Last said" quotes a whole first sentence when it fits. "Last night" stops saying the run is under way once the
    dream is written up. And the wake's first line names the session id, which the reflect tool asks for.
  • What waits on you comes up when it's relevant, not as a pile. An item waiting on your answer is no longer one of
    the wake's still-open lines. Instead, when a turn you start touches one, it's raised once after the recall block,
    with what it needs and how long it has waited; and a session's wake says at most one ("Waiting on , ask
    when it fits: …"), the most relevant to the folder's week first, the same item there at most once a week. Each item is raised at most once a lived day across
    every session, then again after 3, 7 and 14 lived days, and then it's parked (still under recall "still open").
  • "PW" before a version number is no longer read as a password. 0.3.52's prose-password check redacted the
    version in "PW 1.48" (short for Playwright); a bare Playwright version after "PW" is now left alone, and short
    passwords after "PW" are still redacted. The manual and README say "most passwords" named in a sentence.

counterparts 0.3.52

Choose a tag to compare

@mlapeter mlapeter released this 09 Nov 22:26

Counterparts now runs on Windows (new: tell us how it goes), and a password said in a sentence is no longer kept. It also brings memory that fades on purpose as it grows: old per-turn records roll up, older release notes give way to newer ones, finished things stop being overdue, and the night closes items itself. What you'll notice: the morning's "Last night" can say in one line which open items the night closed; doctor has new Growth lines saying what is filling up and whether it leaves as fast, and a Recall steps line saying where recall's time goes; fewer overdue and still-open items at wake. One thing goes for good: the per-turn recall records (recall.decision, recall.credit, recall.rated) are now kept 30 lived days instead of 90, and on the first night after updating, the ones older than that are rolled up into one summary row per day (counts, the memories credited, the sessions) and deleted. If you want them, copy your store before updating. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.50 needs no store restore: counterparts disconnect, 0.3.50's install, then counterparts connect (the rows already rolled up stay gone).

0.3.51 was packed but never published; everything it held is in this entry.

  • Doctor says where recall's time goes. Each recall now records how long each of its steps took (finding words,
    this turn's meaning, the link walk, ranking, the gate, settled pairs, habituation, the render, and the rest), and
    counterparts doctor has a Recall steps line with the last week's median for each, largest first. It only
    measures: what recall shows is unchanged, byte for byte.
  • A password said in a sentence isn't kept. "My wifi password is …", "the PIN is …", "pw for the router
    is …": the value is replaced with [REDACTED:password] and the sentence stays, so the memory still says there is
    one. Ordinary sentences about passwords ("I forgot my password", "the password field is empty") are left alone.
  • The raw conversation text gets the same stripping. What's kept of a conversation until it's written up now
    goes through the same check as a memory before it's saved, on your machine and on hosted. Text saved before this
    version isn't rewritten; it's deleted on its usual schedule. Not every secret has a shape: don't paste things you
    wouldn't want kept.
  • Counterparts runs on Windows. Install it with npm (npm install -g counterparts@latest) and Node 22.15 or newer,
    then counterparts and counterparts connect as on macOS and Linux, from cmd, Git Bash or PowerShell (where
    Windows' default policy refuses npm's scripts, counterparts.cmd works). The hooks connect writes run whether
    Claude Code uses Git Bash or PowerShell for them, and the memory tools register with Claude Code installed either
    way (its installer or npm). It's new and has few users yet: tell us how it goes. Under WSL it works as on Linux.
    QUICKSTART's "On Windows" says what differs.
  • Old routine work notes fade faster once a newer one moves on. When a newer work memory in the same folder
    says a later version in the same series was released or published ("0.3.50 released"), the night fades the
    older status note ("0.3.46 ready to publish", "0.3.47 after-publish steps") the way a revision does: it
    stays findable by id and in facts (folded under the newer one), labelled earlier, drops out of the wake's
    lines, and can be undone. Only routine notes nobody used: anything the owner said,
    anything felt, anything about the owner, us or me, and open items keep their slow fade.
  • A done memory is never overdue. A memory marked status: done whose date is today
    or already passed now reads that date as when it happened: it leaves Coming up (due
    today included), the overdue count and the plain-reminder line, older memories
    included. Writing one through remember or session_end stores the date as
    occurredOn and says so beside the memory. A repeating reminder is unchanged: a done
    there still closes only this time.
  • Old chapters stop turning up as footnotes. A journal chapter whose last entry is two weeks
    (14 lived days) old or more is no longer offered beside a prompt, so its slot goes to another
    memory. It's all still kept: asking (facts or meaning), opening it by id, the wake and the
    night read it as before.
  • Doctor says what is growing and whether it leaves as fast. Five new lines, Growth: memories, event log, cue pairs and dream queue, plus new and swept edges on the
    Association line: how full each is, what came in and what left over the last 7 lived
    days, and which is faster. The memories line also says how many nights the prune takes to
    get round the store, and counts the kinds it never takes (the journal, chapter copies,
    handoffs, cards) apart.
  • The per-turn event rows are kept 30 lived days, not 90. recall.decision,
    recall.credit and recall.rated make up most of the event log. Each day of them is
    summed into one kept row before it goes. The all-time numbers read those sums beside
    the rows still kept, so they don't shrink: the dashboard's counts, counterparts mechanisms, probe-oq4, a memory's days used and the sessions seen. Everything else in
    the log keeps 90 lived days. The table of what each recall turn showed goes after 30 too.
  • The night closes done release steps itself, and stops asking. When a work memory in a folder says a release
    happened ("0.3.50 released"), the night closes the open items in that folder that were earlier steps of the same
    series ("0.3.46 ready to publish", "0.3.47 after-publish steps", "0.3.48 cut"). Each close is recorded with its
    reason and the memory that showed it, and the next wake says what closed in one line under "Last night". A
    question to the owner is never closed this way. What only might be done (a pinned build whose series moved on, a
    PR a later memory says merged, or what the night's reflection wasn't sure of) is no longer asked about in the
    morning share: recall "still open" marks it mayBeDone. Say one isn't done and it reopens (the reflect tool's
    reopen, or a write that updates it and says it's still open).
  • A reminder to do something in one folder is said in that folder. A plain reminder written as another folder's
    work no longer shows whole in this folder's wake, above it, or in the terminal line: it waits for its own folder
    (and any folder joined to it as one project), whose first wake that day says it whole, even when another folder
    woke first. Elsewhere, "in other folders" names it by folder and title (up to two, when there's room) and
    recall "coming up" lists it. Reminders about you, us or me still show in every folder, once, then fold to
    "shown earlier". A claude.ai chat, which has no folder, is told only those personal ones.
  • A deadline in a folder you didn't open still reaches you. A folder's reminder that went unsaid because no
    session ran there is said once, two days you used Counterparts after its day, in whatever session you're in,
    with its folder named ("garden: send the invoice, due Wed 4 Nov 2026"). It never interrupts a prompt, and
    checks you set for yourself wait for their folder.
  • Doctor tells a quiet folder from a missed reminder. A reminder due while its folder had no session is an info
    line, not a warning; it's a warning only if it still wasn't said after that one crossing.
  • An open thread marked "world" stays in its folder. Something the world owes (a reply, a dated check) is in the
    Still open of the folder it was written in, not every folder's. So is a dated item marked world.
  • A write names the open items it looks like it settles, from any folder. When a
    memory you write through remember or session_end says something was done ("0.3.47
    published", "#760 merged"), the result lists up to three open items whose subject it
    matches, wherever they were written, with their folder, and asks you to close any that
    really are settled (updates + unresolved: false). Nothing is closed on its own. An
    item that was a question to the owner says that only their answer closes it.
  • The remember tool no longer promises mood-matched recall. Its feelings text said a
    memory's mood would help find it again; mood has carried no weight in recall since October,
    so the promise is gone.
  • Hosted: a command-line step exits only after its output is written. Under bun, the
    operator's commands (migrate-keys and the rest) could exit before what they printed
    reached a pipe, a Railway log included; they now wait for it (five seconds at most).

counterparts 0.3.50

Choose a tag to compare

@mlapeter mlapeter released this 06 Nov 20:40

Write-only backups for hosted: a leaked environment can no longer open a backup. This release changes only hosted Counterparts (app.counterparts.ai). Its backup keys used to be openable with a key kept in the server's environment; now the server keeps that key on its own disk and puts each backup key in the bucket sealed to a recovery key whose private half is kept offline, so the server can write those keys and never read them back. For every store made from now on, the environment and a copy of the bucket together open nothing. The stores that already existed before the operator's one-time migration keep their keys, so a copy of the bucket taken before it, together with the old environment key, still opens their backups; that is accepted for those stores only, and is why the migration runs before anyone outside the family is invited. The npm package, Claude Code and Claude Desktop behave exactly as in 0.3.49. No store-format change (store v15, cache v7), no migration on your machine. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.49 needs no store restore: counterparts disconnect, 0.3.49's install, then counterparts connect.

  • Hosted: backups are write-only. Each store's backup key, and the key for the account
    and sign-in databases, is put in the bucket sealed to an offline recovery key: the server
    can write it and never read it back. The key the server uses day to day lives on its own
    disk, never in its environment and never in a backup. Restoring after a lost disk takes the
    recovery key, typed in by the operator for that one step.

counterparts 0.3.49

Choose a tag to compare

@mlapeter mlapeter released this 06 Nov 18:15

The hosted beta note, and the app's front door no longer 404s. This release changes only hosted Counterparts (app.counterparts.ai): the account page now says what the beta keeps and how to delete or download it, and the bare address opens your account instead of answering "Not found". The npm package, Claude Code and Claude Desktop behave exactly as in 0.3.48. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.48 needs no store restore: counterparts disconnect, 0.3.48's install, then counterparts connect.

  • Hosted: the account page says what the beta keeps. Right under who you're signed in as,
    one paragraph says what hosted Counterparts stores and what it doesn't, how the disks and
    backups are encrypted, how to delete one memory or the whole account, how to download it,
    and where to send a bug (GitHub issues) or anything else (an email address).
  • Hosted: app.counterparts.ai/ opens your account. The bare address used to answer "Not
    found"; it now goes to the account page, which asks you to sign in first if you aren't.

counterparts 0.3.48

Choose a tag to compare

@mlapeter mlapeter released this 06 Nov 16:37

A lighter tool list for claude.ai, hosted pages that look like counterparts.ai, and a README that explains itself. On claude.ai the tool list a chat loads is about a third smaller; Claude Code and Claude Desktop get exactly the tool list 0.3.47 gave them. The hosted sign-in, Allow, account, delete and error pages take counterparts.ai's look, light or dark as your device is set. The README's first screen, which is also the npm page, says what Counterparts is and what the first run does. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.47 needs no store restore: counterparts disconnect, 0.3.47's install, then counterparts connect.

  • The README opens by saying what it is: long-term memory for Claude Code, written by
    Claude itself.
    Its first screen, which is also the npm page, now leads the install with
    npm (bun on the next line), says what the first run asks and changes, and has a short
    "What it remembers, and where it lives" section, with a link to counterparts.ai's guide to
    Claude Code's own memory. The npm description and keywords name Claude Code too.
  • claude.ai's tool list is about a third smaller. The list a claude.ai chat loads went from
    about 92 KB to 65 KB (roughly 10k fewer tokens on every turn once memory is in use). The
    field descriptions are shorter, and the guarantees that change nothing about how Claude calls
    a tool are no longer spelled out there, though they are still enforced. Words that weren't
    true on claude.ai are gone: a terminal, folders, the end-of-session ask, the dashboard.
    Claude Code and Claude Desktop get the same tool list as before.
  • Hosted pages look like counterparts.ai. The sign-in, Allow, account, pending-delete,
    error and done pages share one inline style in the site's look: its light face by default,
    its dark one when the device is set to dark. No page's words, fields or routes changed, and
    the content security policy is the same.

counterparts 0.3.47

Choose a tag to compare

@mlapeter mlapeter released this 06 Nov 14:20

0.3.47 — 2026-11-04

Folded reminders, honest folders, refused calls kept out of your wake, compacted sessions remembered, and delete one memory from chat (hosted). With the wake's shape on, a reminder a session started today already showed you folds into one line in later wakes that day. A dream files work under the folder it came from, another folder's dated work is counted rather than listed as due today, and other folders go by their names, not their paths. A memory about a refused tool call or a declined permission prompt stays out of the wake and the footnotes, and a morning share no longer asks for a tool call to mark it told. What you say after a conversation compacts is written up at night. "Delete the memory about X" works in chat, on hosted and locally, with 24 hours to undo it. Also: a hosted download opens in a local install and answers straight away, the redactor keeps tokenizer code readable and catches more secret names, and uninstall names the right remover. Existing stores' starter notes about deleting change once, only where they are still exactly as they started. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.46 needs no store restore: counterparts disconnect, 0.3.46's install, then counterparts connect.

  • A reminder you were already shown today folds in later wakes. With the wake's shape on, the first wake of the day
    lists a plain reminder due today whole; a later wake that day, in a session that started today, counts it instead
    (2 due today, shown earlier; recall "coming up"). A compaction's wake folds nothing, and a reminder told only in
    last night's session after midnight stays whole. Dated items that aren't plain reminders show whole as before.
  • The manual's section index says what is inside each section. recall with the handle
    manual now lists every section with a line that names its contents (the phrases, the
    tools, the commands), so looking for how to pause or how reminders work no longer means
    guessing from a generic opening sentence.
  • On claude.ai, the first chat of the day runs the night. Before answering you, Claude
    sorts the day before (day's end, a dream, a reflection, about half a minute), then
    answers in the same reply and says "I sorted yesterday's memories just now. Tell me if
    you'd rather I ask first next time." One chat at a time is told (a later one only if the first never
    ran it), and none once a night has run, whether from another chat or your nightly routine. Say "ask first" to have it
    ask before it starts, or "no dreams" to turn nights off. A nightly routine still works,
    once Counterparts' tools are set to Always allow (the routine's form lists only those).
  • The night names its conversation on every step. The dream's prompt now passes the
    session on its propose and journal steps too, so with two chats open a night's changes
    stay with the chat that runs it.
  • Tokenizer code is no longer redacted. Code lines that assign one identifier to another
    whose last word is token, key or id (pad_token set to eos_token; letters only, no digit,
    lowercase name, =) are kept as written; env vars, YAML, passphrase-shaped
    values, quoted values, digits and passwords still redact as before.
  • More credential names are redacted. passphrase, db_pass and other *_pass names,
    a whole-name pin with a numeric value, private_key, signing_key, session_key
    and a bare capitalised AUTH= / BEARER= now keep their name and lose their value;
    pin_memory=True, session_key_fn and key-file paths such as private_key = "keys/a.pem"
    stay readable.
  • Prefixed PINs and suffixed AUTH are redacted too. user_pin, SIM_PIN, PIN_CODE,
    pinCode, gpgPassphrase and BASIC_AUTH= keep their name and lose their value.
  • Ask Claude to delete one memory, with a day to change your mind. On hosted and on a local
    install, "delete the memory about X" makes Claude find it, say which one it is, and delete it:
    it's hidden at once from recall, the wake, footnotes and the night, and erased for good after
    24 hours. "Undo that" brings it back whole before then; recall's handle deleted lists what's
    waiting, and the wake counts it. On hosted the erase runs on the server's own timer and the
    backups let go of it within about a day after (the hosting provider's own short-term disk
    snapshots rotate on its schedule); on a local install it's erased at the first
    session after the 24 hours. counterparts remove <id> --confirm still deletes at once, and
    finishes a memory deleted in chat.
  • Recalling a removed memory by its id no longer fails with an internal error; it reads as
    not found.
  • The starter notes say so too. Two of them changed with this: the one about seeing and
    changing what Claude keeps now says one memory can be deleted in chat and undone within a day
    (on hosted it used to say a memory couldn't be removed yet), and the tips list gains "Delete
    that". On hosted, the note on how memory works gets the new night words too (the day's first
    chat runs the night). An existing install gets the new words once, at its next session, but only
    where those notes are still exactly as they started; one you corrected or let go stays as it is.
  • Uninstall names the right remover and cleans up after itself. Its last line says
    npm uninstall -g counterparts for an npm install and bun remove -g counterparts for a
    bun one (both when it can't tell), where it used to say the bun line to everyone. When it
    takes its hooks out of ~/.claude/settings.json, it also removes the backups of that file
    that connecting and disconnecting kept, and says so when the file is left as {}. When
    there were no hooks of ours in the file, the backups stay and it says where.
  • The docs say what goes over the network. README, the quickstart and setup now say the
    one request Counterparts makes of its own: a daily check of registry.npmjs.org for the
    newest version number, sending nothing about you, and how to stop it ("updateCheck": false in the configuration, or asking Claude to stop the update reminders). The export example now
    works as shown (--out with --passphrase or --plaintext), and the docs say no command
    opens an encrypted export yet.
  • Work stays in its own folder, and folders go by their names. A dream now files what it writes
    under the folder its memories came from, so a summary of one repo's work no longer shows up as
    another repo's news. Work drawn from several folders goes to the one most of it came from;
    something personal drawn from several folders belongs to none, so it reaches them all. Another folder's dated work
    is no longer listed as "Due today" in this folder's wake: it is counted under "in other folders"
    (personal reminders still reach every folder). And the wake names other folders as sim or
    paper instead of their full paths, falling back to the path only when two folders share a name.
  • A morning share no longer asks for a tool call to mark it told. When last night's share
    reaches a session, handing it over marks it told, so a new person isn't shown a permission
    prompt for bookkeeping they have no context for. A session that still calls reflect with
    told gets a plain yes.
  • A memory about Counterparts' own refused calls stays out of your wake. When a session
    writes down that a tool call was denied, that a permission prompt was declined, or what it
    couldn't do (a slash command, a tool it doesn't have), the memory is kept without a date, an
    open thread or a "me", "us" or owner mark, so it never leads Coming up, the wake's news or
    a prompt's footnotes; asking for it still finds it. The person's own words about a prompt
    are kept as usual.
  • What you say after a conversation compacts is written up, even when the session had
    already saved its own notes.
    A session that wrote its handoff or chapter, then compacted,
    then carried on for a few quick lines used to have those lines treated as a short closing
    tail and never written up. Now anything said after a compaction counts as not yet
    recorded, so the next night writes it up, unless the session records itself again after
    the compaction.
  • Hosted: /readyz says when backups stop reaching the bucket. The hosted server now checks
    every 10 minutes that its backups are actually landing in storage, not only that the backup
    process runs. After 30 minutes with nothing landing, /readyz answers 503 with
    backups: "stale", and the log says so with a backup.uploads.stale line. A fresh start gets
    those 30 minutes before it can read stale.
  • Hosted: invite codes now expire after 7 days. A code that leaks later can't sign anyone up. Codes made before
    this change keep working until they're used. Every page from app.counterparts.ai now tells your browser to use
    https only, and the server no longer advertises two sign-in features it never used (DPoP and the library's own
    sign-out page).
  • A hosted "Download my memory" file opens in a local install and answers questions straight
    away.
    The download carries only the memory file, not its search index, so a question used to
    find nothing until counterparts verify --rebuild was run by hand. Now the first session that
    opens a memory with an empty search index rebuilds it, whichever way it opens (Claude Code,
    Claude Desktop, the console, the background worker). Search by meaning fills in as usual.
    counterparts doctor shows a "Word index" line when the index is empty (amber) and for a week
    after an open rebuilt it, and counterparts ask, which only looks, says why it found nothing.
  • Starter notes follow the memory to where it's opened. A hosted memory opened by a local
    install gets the local wording of its starter notes (term...
Read more

counterparts 0.3.46

Choose a tag to compare

@mlapeter mlapeter released this 04 Nov 16:13

New stores get starter memories and the new wake layout; existing stores barely change. A store counterparts install (or start-fresh, or hosted signup) creates begins with eleven starter notes, "told, not lived", counted apart from your own memories, and its sessions start in the wake's shape, which now opens with what happened since the last night, steps an old handoff down to one line, drops empty sections and settles dated things that already happened. A store that already exists gets no starter notes and keeps its wake-shape setting; where the shape is on (counterparts wake-shape on), its wake gains those same changes, and on any store a reminder that a newer done memory says already happened is no longer told as due. Also: write-ups leave out permission prompts and refused tool calls, the first meeting asks easier questions, a manual ships with every install (recall the handle manual), a memory called noise twice stays quieter for 14 lived days rather than 7, and the plugin downloads its program once. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.45 needs no store restore: counterparts disconnect, 0.3.45's install, then counterparts connect.

  • A memory two sessions called noise now stays quieter for two weeks, not one. Rated 0 in two sessions within 14 lived
    days (it was 7), it faces the same raised bar until those ratings age out or a later 2 or 3 clears it. A turn that
    reaches it strongly still shows it. Right after installing, the longer window starts counting at the next rating.
    recall.HABITUATION_WINDOW_DAYS back to 7 undoes it.
  • The plugin downloads its program once on Ubuntu 26.04. On a computer with no Bun or Node,
    the hook and the memory server starting together could both download the program (about
    56 MB), because that Ubuntu's mkdir can tell two processes they each made the same
    directory. The launcher now takes its lock with the shell itself. The same double download
    could follow a download that died, on any system, and now doesn't. Nothing unchecked ever
    ran: each download was checked and moved into place in one step. And on a Mac whose sh
    is zsh, a first start now downloads the program at all: it used to say "getting ready"
    and never start.
  • Hosted: Litestream's own log lines reach the server's log. A backup that fails to sync (a refused upload, an
    R2 permission error, a rejected encryption key) now says so in the log, and a redeploy's log shows Litestream's
    final sync finishing. Before, those lines were dropped and only the process being up was visible.
  • Hosted: the restore drill is in the repository, and the KEK check is by fingerprint. hosted/drill/drill.sh
    restores production's deployed image from a drill bucket on R2 with a throwaway key, and writes its results outside
    the repository. The offline copies of the key are now checked by a short fingerprint against Railway's, so the real
    key no longer goes into a drill. The runbook says what to do when a restore refuses a mistyped key.
  • Recall's per-conversation bookkeeping no longer piles up. The record of what each
    conversation was already shown (so nothing repeats) is now cleared once a conversation
    has gone unused for the retention window, 90 days of use by default, in the nightly run
    beside the old-version cleanup. A conversation used inside the window keeps all of it.
  • A new store starts with eleven starter memories, "told, not lived". counterparts install (and
    start-fresh) and hosted signup plant a short set about how this memory works and how to use it well, in each
    surface's own words, marked with the source told and the set's version in the store's meta. They start weak,
    fade unless used, and can be corrected or removed like any memory. Recall fills its footnotes with every lived
    memory first and never shows a starter one loud; the deliberate recall ranks one below a lived match of
    comparable score. The night, the wake, the core, beliefs and cards never read them. An upgraded store never gets
    them. No store-format change.
    status, doctor and the dashboard count them apart ("0 memories · 11 starter notes"), and uninstall says how
    many of the memories it counts are starter notes. init still makes a bare store; the harness week starts with
    the set unless given --no-told.
  • Nightly write-ups leave out permission prompts and refused tool calls. When you decline a
    permission prompt, the host's "interrupted for tool use" line no longer reaches the write-up
    of that conversation, and whoever writes a session up (the nightly catch-up, or the next
    session pointed at it) is told that permission prompts, refused tool calls and the
    memory system's own notices aren't things that happened between you, unless you talked
    about them. A night had saved "declined permission to mark a note delivered" as a memory.
  • The first meeting asks four questions now, and easier ones. After your week, it asks "What are you working
    on right now?", then what you've been into lately that has nothing to do with work, then a good Saturday. "What do
    you love most about what you do?" is gone, and so is "What got you started building things?". A meeting already
    under way when you update finishes as it started: the answers it gives under the old questions are still kept.
  • Starter memories no longer push out a lived memory. On a new store, a seed that shared a word with the question
    set the deliberate recall's "weak match" bar, and a lived memory found by meaning alone fell under it ("which
    coffee maker do we own?" lost the note about the espresso machine). Facts mode now lists every lived match before
    any starter one, and judges a lived match only against lived ones. The turn's footnotes read their floors from the
    lived memories too, so eleven starter notes no longer double the bar a young store's memories must clear (a
    starter memory still faces the stricter whole-store bar, so a loosely matched seed is not footnoted on a turn nothing lived answers). The
    single-binary check now runs on every pull request that touches src/.
  • The counterpart has a manual. docs/manual.md says how Counterparts works and how to
    use it: the everyday phrases, folders, nights and dreams, seeing and changing what's kept,
    export, backup and uninstall, checking health, and how hosted differs. It ships with every
    install, so it matches the version you run and works offline. Your counterpart opens it
    with recall's handle manual (a list of sections), and manual <section> opens one.
  • The wake knows yesterday's news. In the wake's shape (on for new installs, below), a session start opens, after
    the page, with "Since the last night:": up to three of the strongest things about you, or about the two of you, saved since the last
    night ran, from any folder that's on. Something said in another folder this morning (a launch, good news from the vet)
    no longer waits a night to reach the next session. Nothing new, no section.
  • An old handoff stops leading the wake (the wake's shape, as the next one). Half a day after a folder's last handoff, once another session has worked
    there and left something, the handoff shows as one line with its age and how many sessions came since, after what they
    did, instead of whole and first for days. A quick check-in that saved nothing, or a session still open beside yours,
    leaves it whole.
  • No empty sections in the wake. "Nothing open.", "Nothing dated." and "Open here now: no other session." are gone,
    with their headings: a section with nothing in it isn't printed.
  • Something that already happened stops being "due today". When a newer memory marked done says a dated thing
    happened (the shop launched a day early), its reminder isn't told and it leaves Coming up. Only on or after its day,
    and only by a memory from the day before it or later, so last week's vet visit never settles this week's.
  • A new install starts with the wake's shape on. The top-down wake (where this folder stands, elsewhere, coming up,
    still open, yesterday, last night, each said once) is now how a new store's sessions start. A store that already existed
    keeps its setting (counterparts wake-shape on|off|default).
  • Help no longer says export is encrypted by default. It never was: export refuses until you pass
    --passphrase (encrypted) or --plaintext (said on purpose). counterparts --help, counterparts export --help,
    the README and the quickstart now say so.
  • On claude.ai, the scope tool speaks of one memory, not folders. Its description, its read and the refusal
    while memory is off or paused say one account is one memory, to pause, turn off and turn back on; they no longer
    show the server's file path or a counterparts scope command nobody on claude.ai can run. Claude Code's and
    Claude Desktop's words are unchanged.
  • A new store's starter note on using memory well points to the manual (recall the handle "manual"). Stores
    that already exist keep their notes as they are.

counterparts 0.3.45

Choose a tag to compare

@mlapeter mlapeter released this 02 Nov 02:44

Nearly nothing you'd notice changes. One thing happens at once: the first pass after upgrading deletes the raw copies of notes saved with no session or with counterparts note that are already past their age (3 days once the memory has landed, 14 whatever). Memories aren't touched, and raw text deleted this way doesn't come back, even if you go back to 0.3.44. Learned cue words skip path fragments and acknowledgements. A memory that couldn't be saved can be sent again. A laptop's sleep no longer stops the night's work, and a headless session leaves the night to the next one a person starts. No store-format change (store v15, cache v7), no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.44 needs no store restore: counterparts disconnect, 0.3.44's install, then counterparts connect.

  • Learned cue words skip path fragments and acknowledgements. Words from file and socket paths (tmp, uds, ...),
    hex ids, and replies like "sounds good, go ahead" no longer wire themselves to a memory, and what they learned before
    is ignored. Nothing in your store is changed.
  • Notes saved with no session, or with counterparts note, now lose their raw copies on time. Those copies shared
    a label, so each new note restarted the clock for the old ones, and with a note every couple of days the oldest was
    never deleted. Each now goes on its own age: 3 days once its memory has landed, never more than 14. The memories
    themselves aren't touched. The first run after updating deletes every such copy already past those ages, at once;
    doctor's Raw transcripts line says how many.
  • A laptop's sleep no longer stops the night's work. A night's watchdogs count time awake now, so a machine that
    slept mid-run picks up where it left off instead of killing the child on waking. What a run slept through is on its
    row, and a run that's asleep isn't read as lost, so no second one starts beside it. A child that's really stuck is
    still stopped on schedule.
  • A headless session no longer starts the night. claude -p and the Agent SDK leave it for the next session a
    person starts, which runs it as usual, so nothing is lost by waiting. counterparts catch-up is run by hand as before.
    A store that only headless sessions use gets no nights, and doctor says so after two days.
  • A memory that couldn't be saved can be saved again. When the store was locked or full
    at the moment of writing, the note was recorded as handled anyway: nothing was stored, sending
    it again was refused as a duplicate, and its raw copy started its 3-day clock, along with
    everything else that session had said if it was a session_end. Now nothing is marked
    until the memory is on disk, so sending it again saves it once, and an unsaved note's words
    stay owed for a later write-up. remember says what happened and to send it again, and the
    failure shows in doctor. A write-up part with a memory that didn't land stays open to be
    sent again, unless it's the last part and another memory from it landed; then the reply
    says to save the rest with remember.

counterparts 0.3.44

Choose a tag to compare

@mlapeter mlapeter released this 01 Nov 23:47

0.3.44 — 2026-11-01

Nothing you'd notice changes by default. counterparts page-form own is new: one store's nightly writers amend its self page in the page's own form, and it's off until you set it (counterparts page-form default puts it back). The core door's shadow has stopped recording; what becomes core is decided exactly as before and nothing in it moves. No store-format change (store v15, cache v7); the switch is one setting kept in the store, with no migration. Nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. Going back to 0.3.43 needs no store restore: counterparts disconnect, 0.3.43's install, then counterparts connect; 0.3.43 ignores the switch.

  • A store can keep its self page in its own form. counterparts page-form own tells the nightly writers to amend
    the page keeping its sections and headings, dated stories where it tells stories, instead of keeping "## How I
    work" as a few short lessons. Off unless set, one store at a time; counterparts page-form default puts it back.
  • The new way of deciding the core no longer runs in shadow. 0.3.36 started it beside the old one, recording each
    night what it would have let in or kept out. Those records have stopped. The core is still decided exactly as it was,
    and nothing already in it changes. The code stays in place, and physics.CORE_DOOR_SUPPORT set to "shadow" turns
    the records back on.