ByteRover — durable project memory

ByteRover keeps project knowledge as a deterministic context tree. You are the brain; the engine never calls an LLM. You do the reasoning, then run the bundled deterministic scripts (via your Bash tool) to retrieve or persist knowledge.

Iron Law

1. Query before you think. At the start of any non-trivial task, retrieve

what the project already knows so you don't re-derive or contradict it.

1. Curate after you implement. When you finish something worth remembering

(a decision, a gotcha, a pattern), write it back so the next session inherits it.

Where memory lives

The engine resolves the right context tree from your working directory. You do not compute, look up, or override the storage path. Run commands from inside the project (any subdirectory works) and the right tree is found automatically — the resolver and registry are engine-internal.

Language

Match the user's input language. Author the human-readable content in whatever language the user is conversing in — body text of <bv-*> elements, list items, and the title / summary attributes on <bv-topic>. There is no setting and no flag; you infer the language from the conversation, the same way you already do when replying.

The schema stays English. Keep tag names (<bv-fact>, never a translated tag), attribute names (subject, category, severity), enum values (category="project", subject="visa_required"), and the path attribute in English for tooling consistency. Code snippets and identifiers stay verbatim.

Concrete (Vietnamese conversation): path="travel/vietnamvisa" and subject="visarequired" category="project" stay English, while title="Yêu cầu visa Việt Nam" and the <bv-fact> body (Người Mỹ cần visa để vào Việt Nam.) are Vietnamese. Same rule for Chinese, Japanese, Korean, Russian, Arabic, etc.

Commands

Run from inside the project (any subdirectory). Topics are stored as <bv-*> HTML. Spaces (including the default) are provisioned through the ByteRover desktop app — the agent never creates them.

Required <bv-topic> attributes: path and title MUST be present and
non-empty on the topic root or the writer rejects the call. The path on
<bv-topic> should match the positional <topic-path> you pass to
record.mjs. See [vocabulary.md](vocabulary.md) for the full attribute spec.

| Read | node scripts/brv.mjs read "<topic-path>.html" [--raw] (use --raw to get the on-disk HTML byte-for-byte — required when round-tripping through record --html, since elements[].text flattens inner <ul>/<li>/<strong> formatting) | | List | node scripts/brv.mjs list | | Rebuild manifest + index | node scripts/manifest.mjs | | Detect consolidation opportunities | node scripts/dream.mjs --mode merge\|link\|prune\|synthesize (proposes — never executes) | | Fold one topic into another | node scripts/merge.mjs <survivor> <loser> [--title "…"] [--summary "…"] [--tags a,b] [--keywords a,b] [--reason "…"] [--related @x.html,@y.html] (use --title to give the survivor a new title that reflects the merged scope — without it, the survivor keeps its own title even when the merged content is broader) | | Remove redundant topics | node scripts/prune.mjs <topic-path>... (variadic, atomic — aborts on first failure with a partial report) | | Combine many topics into a new one | node scripts/synthesize.mjs <new-path> --html '<bv-topic …>…</bv-topic>' --absorb a.html,b.html,c.html | | Rename / move / re-parent a topic within the current space | node scripts/move.mjs <topic-path> --to-path <new-path> (atomic — siblings' related= refs are rewritten to the new path; signal sidecar carries over; covers all of rename, move-to-new-parent, deep re-parent, and combined) | | Add or remove a related= cross-link between two topics | node scripts/brv.mjs link "<a>" "<b>" [--bidirectional] [--remove] (atomic byte-preserving edit of the <bv-topic related="…"> attribute — use this instead of hand-editing the opening tag, which is unsafe for summaries containing > characters; idempotent — re-linking is a noop; refuses to add a ref to a topic that doesn't exist (would dangle) — but --remove allows missing targets so you can clean up existing dangling refs) | | Move a topic to a different space | node scripts/move.mjs <topic-path> --from-space <name> --to-space <name> (rebuilds manifest + index in both spaces; reports any sibling related= refs in the source that now dangle) | | Migrate legacy markdown | node scripts/migrate.mjs [--dry-run] | | Migrate a self-hosted v3 VPS → v4 (full guided procedure) | follow [migrate-v3-to-v4.md](migrate-v3-to-v4.md) | | Authenticate / log in to ByteRover | node scripts/auth.mjs — returns at once; the user approves in the browser, then auth.mjs status; see [Authenticate with ByteRover](#authenticate-with-byterover) | | Check whether signed in (local, no network) | node scripts/auth.mjs whoami — prints { ok, authed, providerKind } (a credential exists locally; not a server-side validity check) | | Log out of ByteRover | node scripts/logout.mjs | | See which space this folder uses | node scripts/space.mjs current | | List spaces (grouped by team_name) | node scripts/space.mjs list | | Bind this folder to a space | node scripts/space.mjs bind "<name>" |

To delete a topic, surface the request to the user — deletions happen through the ByteRover desktop app.

Every command prints a JSON result. On { "ok": false, ... } read [troubleshooting.md](troubleshooting.md).

Citing retrieved memory

After a query answer materially drew on the returned hits, append the citation_block from the query envelope to your response so the user can see — and click through to — what was used. The engine pre-formats the block; your job is decisional, not formatting.

When the query envelope has should_cite: true AND your answer genuinely used the retrieved memory:

1. Add ONE blank line after your answer.
2. Emit the citation_block string **verbatim — character for

character** — including the […](…) Markdown link wrapping each title, the indented "snippet" line when present, and the Updated by <agent> · <relative> trailer when present. Every host the engine targets (Codex CLI, Claude Code CLI, Claude Code Desktop) renders the wrapped title as a clickable inline link.

1. Don't reformat, paraphrase, or add agent-generated relevance scores

to the citations. The user judges relevance from the title.

1. Do NOT unwrap the Markdown link. The engine deliberately emits

- Title · Recalled N× — rewriting it as a plain title with the URL on a separate row undoes the format that hides the long URL behind clickable text. Same rule for the snippet and updated-by lines: keep them indented exactly as the engine emitted them, never merged into the title row or stripped of their leading spaces.

When should_cite is false, OR when you didn't actually use the returned hits to answer, OR when the user asked for no sources — skip the block entirely. The engine intentionally suppresses low-confidence retrievals; don't second-guess by surfacing them in prose.

Example final response shape (the 📚 … block comes from the engine, emit it verbatim):

ATM: Er Rak Error is a 2012 Thai romantic comedy by Mez Tharatorn …

📚 From ByteRover:
- [ATM: Er Rak Error (2012 Thai film) (fact)  ·  Recalled 14×](http://127.0.0.1:<port>/i/topic/<space-id>/context-tree/movies/atm_er_rak_error.html#bve-a1b2c3d4)
    "matched-context preview that explains why this hit ranked here"
    Updated by claude · 3d ago
- [Pee Mak Phra Khanong (2013) (highlights)](http://127.0.0.1:<port>/i/topic/<space-id>/context-tree/movies/pee_mak_phra_khanong.html#bve-e5f6g7h8,k1m2n3p4)

Authenticate with ByteRover

Run this when the user wants to authenticate / log in / sign in / connect / link ByteRover for this workspace (or when ByteRover reports the session has expired):

node scripts/auth.mjs

This returns immediately with { "ok": true, "pending": true, "verifyUrl": …, "userCode": …, "expiresInS": … } while a background process waits for the approval. Relay the verifyUrl (and the code) to the user verbatim, tell them how long the code lives (expiresInS, typically a few minutes), and tell them to come back and say "approved" once they have approved in the browser. End your turn — do NOT poll or re-run commands while waiting.

When the user returns, confirm with:

node scripts/auth.mjs status

• { "state": "approved" } → connected; tell the user.
• { "state": "pending" } → not through yet — ask them to finish the browser

step (the result re-includes verifyUrl) and say so again.

• { "state": "expired" } → the code died; run node scripts/auth.mjs again

for a fresh one and re-relay it.

• { "state": "denied" } → the user declined; stop and surface it.

Re-running auth.mjs while a flow is pending is safe — it returns the SAME code ("reused": true), never a competing one. No secret is ever pasted into the chat — only the short code. To log out / sign out:

node scripts/logout.mjs

For non-interactive / CI environments only, an API key may be supplied by file (never inline in the conversation): node scripts/auth.mjs --key-file <path>. A blocking variant that waits in the foreground until approval (for scripted use, NOT for conversations) is node scripts/auth.mjs --wait.

Authentication is not required — every command works without it. See [sync.md](sync.md) for what authenticating enables, and [troubleshooting.md](troubleshooting.md) if auth.mjs returns { "ok": false }.

Authoring rich topics

Topics render in four sections and you should author with that structure in mind. Each <bv-*> element belongs to one section; the renderer ignores order within a section but you should emit them in section order to keep the source readable.

Every topic MUST include:

1. A <bv-reason> explaining the WHY of the curation (the topic isn't

useful without context for why it was captured).

1. A <bv-task> describing what the topic is about (one sentence), OR

an <h1> + intro <p> if "task" doesn't fit.

1. At least one Narrative or Raw Concept element beyond the task:

<bv-changes> / <bv-files> / <bv-flow> / <bv-structure> / <bv-dependencies> / <bv-highlights> / <bv-decision> / <bv-bug> / <bv-fix>. A topic with only <bv-fact> siblings is a placeholder, not a topic.

1. <bv-timestamp> in ISO 8601 if the content has a reference date.
2. **`<bv-fact subject="snake_case" category="…" value="…">canonical

statement</bv-fact> for each extracted fact — subject is snake_case, category is one of personal / project / preference / convention / team / environment / other, value` carries the extracted form, and the body text is the canonical natural-language statement** (not a label).

Worked example — research/profile-style topic

This is the shape an agent must author. Copy this skeleton and fill the slots; do not invent a flatter shape:

node scripts/record.mjs "movies/atm_er_rak_error" --html '
<bv-topic path="movies/atm_er_rak_error"
          title="ATM: Er Rak Error (2012 Thai film)"
          summary="GTH 2012 romantic comedy by Mez Tharatorn — secretly-dating bank employees bet their jobs on recovering money from a malfunctioning ATM."
          keywords="atm_er_rak_error,gth,thai_romcom,mez_tharatorn,2012"
          tags="film,thai,gth,romcom"
          related="@movies/pee_mak_phra_khanong.html">

  <bv-reason>
    Anchor topic for cross-references between Thai GTH romcoms; documents
    the studio box-office record holder before Pee Mak (2013).
  </bv-reason>

  <bv-task>Profile of ATM: Er Rak Error (2012) — credits, plot, awards, legacy.</bv-task>

  <bv-timestamp>2012-01-19</bv-timestamp>

  <bv-highlights>
    Premise: a bank no-fraternization policy forces a secret 5-year couple
    into a winner-keeps-the-job bet over money from a glitching ATM.
    Was GTH biggest hit until Pee Mak (2013) topped it.
  </bv-highlights>

  <bv-structure>
    <p>Credits:</p>
    <ul>
      <li><strong>Director:</strong> Mez Tharatorn</li>
      <li><strong>Writers:</strong> Mez Tharatorn, Aummaraporn Phandintong</li>
      <li><strong>Studio:</strong> GTH (GMM Tai Hub)</li>
    </ul>
    <p>Lead cast:</p>
    <ul>
      <li><strong>Sua:</strong> Ter Chantavit Dhanasevi</li>
      <li><strong>Jib:</strong> Ice Preechaya Pongthananikorn</li>
    </ul>
  </bv-structure>

  <bv-flow>
    Sua and Jib are bank staff who dated secretly for 5 years under the
    bank no-fraternization policy. An ATM glitches and dispenses 130,000
    baht. They bet: whoever recovers the money keeps their job. The
    sequence escalates into a chaotic showdown involving manager Pakorn,
    Sergeant Sam, and Sam pet crocodile Jack.
  </bv-flow>

  <bv-fact subject="release_date" category="project" value="2012-01-19">
    The film was released theatrically in Thailand on January 19, 2012.
  </bv-fact>

  <bv-fact subject="runtime_minutes" category="project" value="123">
    The film runs 123 minutes.
  </bv-fact>

  <bv-fact subject="box_office" category="project" value="150.11 million baht">
    The film grossed 150.11 million baht, reaching Thailand 7th highest-
    grossing film within four weeks of release.
  </bv-fact>

  <bv-fact subject="record_vs_pee_mak" category="project" value="GTH highest-grossing film until Pee Mak (2013)">
    ATM was GTH/GMM Tai Hub highest-grossing film until Pee Mak surpassed it in 2013.
  </bv-fact>

  <bv-changes>
    <li>2013-11-30 to 2014-02-08: TV sequel ATM 2: Koo ver Error Er Rak aired (21 episodes, One 31).</li>
    <li>Chinese-language remake released as Welcome to the Beartown.</li>
  </bv-changes>

</bv-topic>'

Note what carries WHAT:

• <bv-reason> carries the WHY of this curation. Missing this is the

single most common authoring failure — don't skip it.

• <bv-task> + <bv-timestamp> carry the WHAT and WHEN.
• <bv-highlights> + <bv-flow> carry the HOW (the explanation,

the process).

• <bv-structure> carries grouped lists like cast/credits/file

layouts — NOT a pile of <bv-fact> siblings. Reserve <bv-fact> for discrete queryable claims (release date, runtime, box office, awards).

Output contract (hard rules)

The schema is permissive — the writer accepts violations silently — but output that ignores these rules ranks worse and reads worse:

• All attribute values are double-quoted strings. Not single quotes.
• All attribute names are lowercase.
• Path segments are snake_case with underscores between words:

movies/atmerrak_error, NOT movies/atm-er-rak-error (hyphens).

• related= uses @path.html for file targets, @path bare

for folder/domain targets. The @ prefix is required.

• No code fence wrapper. The --html argument is bare HTML — first

character <, last characters </bv-topic>. No ` , no `html .

• **Don't set createdat, updatedat, id, importance, maturity,

or recency** on <bv-topic>. The writer system-stamps the first three; the last three are sidecar runtime signals the LLM does not author.

• Don't invent custom elements outside the 19-element <bv-*>

vocabulary, or attributes outside each element's documented schema.

Sensitivity — mark facts you intend to share

A topic can be shared at three views: full (everything), redacted (only <bv-fact disclosure="public"> plus structural prose), and metadata (an opaque catalog handle — no content). Mark a fact you intend to share with disclosure="public"; otherwise the fact defaults to restricted and is stripped from the redacted view.

Three rules to write by:

• <bv-fact> is the SOLE unit of per-item restriction. The topic

title and all prose text (inside <bv-structure>, <p>, <bv-flow>, <bv-highlights>, etc.) are public-by-contract — they survive the redacted view verbatim. Never put a secret in the title or in prose; move it into a <bv-fact> (which can then default to restricted).

• Absent or misspelled disclosure is treated as restricted. Fail-

closed: a typo never accidentally publishes, but it can accidentally hide. Double-check the spelling.

• <bv-topic visibility> does NOT make facts public. It's an

informational label only. The redacted view consults each fact's own disclosure attribute, never the topic-level default.

See [vocabulary.md](vocabulary.md#sensitivity-and-the-disclosure-model) for the full model, the canonical example, and guidance on choosing public vs restricted for any given fact.

Two record forms

• Simple form (--title --body) wraps the body in a single

<bv-fact>. Use it ONLY when the topic is genuinely one fact ("we use OAuth 2.0 with PKCE"). For anything else the wrapper structure is wrong.

• Rich form (--html) is everything else — profile dumps,

decisions+reasons+rules+facts, bug+fix runbooks, research notes.

For the full element registry, schema-level attributes, and the authoritative anti-patterns list, read [vocabulary.md](vocabulary.md). For commit-quality recording guidance, read [record.md](record.md).

When to read what

• First time in a repo, or no memory exists for the project → [onboarding.md](onboarding.md)
• Before retrieving → [query.md](query.md)
• Before saving knowledge → [record.md](record.md)
• Choosing the right <bv-*> element → [vocabulary.md](vocabulary.md)
• Periodic cleanup / consolidation → [dream.md](dream.md)

Note: topics are <bv-> HTML (not markdown). scripts/.mjs are generated
from @byterover/skill-runtime by pnpm build:skill; if missing, build first.