Release by release

Memory answers in 1.2 s where it took 3.2 s

Improvement

A B2B distributor with 6,000 documents waited 3.2 seconds, on average, for its context answer. On bad days, 5 seconds. We turned on it the retrieval path that already runs in our own workspace. After 24 hours, the median was 1.2 seconds and the worst case 1.8. Each answer brought 9.4 documents instead of 7.9, and no call failed. The comparison uses real calls: 18 after the change against 140 from the previous eight days.

The right context shows up more often, with fewer documents

Improvement

We measured search against the real corpus of two accounts, asking each document about its own subject. One in four did not show up even among the first ten. Two causes. The filter that avoids repeating a similar document treated two documents from the same client as if they were the same memory, and pushed to the end of the queue the one that answered; it now only demotes an earlier version of the same document or a near-identical copy. The other cause was the cap on documents per answer, too low for a large corpus: with 8 instead of 6, the chance that the right document shows up among the first ten rises 7 points in a 6,000-document account. The delivered context is about 30% smaller than it was with 12.

Improvement

In an account with 580 recorded agent sessions, each session became dozens of documents and the whole conversation stayed indexed. That was 3,711 documents carrying text nobody queries. Each session now leaves a 36-byte anchor with the same identifier, and the text stays on disk where it came from. The corpus of that account went from 38,600 to 34,900 documents, 1.1 million orphan vectors were removed and the indexes are 53% smaller. What was extracted from the conversation (episodes, decisions, facts) does not change and still points to the anchor.

Improved

Old titles used internal jargon (benchmark codenames, acronyms, correction jokes). Each entry now says in one sentence what changed for you.

Fixed

Public pages describe exactly what the product ships today: structured memory per domain, with extra per-area types as enterprise customization.

Fixed

Each vertical shows the real mechanism — hooks capture the area's decisions, handoffs, problems and lessons — and extra per-artifact types are clearly positioned as enterprise customization.

Fixed

Memory browse repeated documents stored in collection sub-folders — 1 in 3 rows was a duplicate on the largest tenant. Every document now appears once, and documents over 2 KB are no longer hidden.

New feature

Sign-in, sign-up and agent-linking screens now match the site's look. Anyone coming from a plan sees the pick confirmed (Solo, Core, Operations) and starts on Free, no card.

Security

• Registration no longer accepts companyId/role; tenant and role selection happens only through an authenticated invite.
• Legacy JWT sessions are revalidated against an existing, active, non-deleted user; identity headers are derived from the credential, not from claims.
• Company offboarding revokes tenant tokens, suspends the tenant in the Engine and then deactivates local data in one transaction.
• OpenBao Agents now deliver the SaaS secrets; unit files and tracked examples no longer carry credentials.

Fixed

• audit_logs test teardown disables the chain trigger as well as the row guard, so the append-only contract stays meaningful in tests.
• Two orphan i18n keys were removed.
• Test database provisioning grants the app role the schema privileges it has in production.

Validation

• Full unit suite green on the isolated CI database as the ci role: 896 pass, 0 fail. Typecheck and lint are green (lint: 0 errors).

Improved

Buttons, fields and selectors have clearer focus in light and dark themes. Green links and field borders also have stronger contrast.

Fix

When the answer-generation service is down, the chat now replies in-flow with an honest notice that the answer could not be completed right now and asks you to try again shortly, instead of a total-failure box. Usage limits and quota refusals keep their own messages as before.

Feature

Agents can discover ValorBrain from the public site, read all 72 public pages as Markdown, and find the MCP server, OAuth, OpenAPI, credential rules, and three hash-verified skills. Setup keeps each credential type separate and no longer suggests unavailable commands or puts secrets in shell history.

Connect an agent →

Improved

We swapped the reader to GLM-5.3 Flash (then identified as Ox Alpha) and turned reasoning effort to max: 69.3% to 75.5% average score on the 400-question memory benchmark (323/400 correct, 80.8% binary). Result setup: reader GLM-5.3 Flash (Ox Alpha); judge GLM-5.2; effort=max; run 2026-08-23; ~44% slower than default effort; this benchmark setup is not the chat default. Along the way we moved dense HNSW indexes from float32 to halfvec (3.6× less disk, 4.7× faster, recall intact), fixed an O(n²) validation that kept index builds at hours, and built, then disabled, a ColBERT rerank that helped synthesis but hurt abstention and extraction.

Read the full report →

Fix

Second wave of the launch audit. The daily value email had answered Unauthorized on all 54 runs since it shipped (a literal %s inside a systemd command) and had never delivered anything; it now goes out at 07:00 in the workspace timezone. In per-user workspaces, sync always ran as the first member: their token re-synced every 15 minutes while everyone else's never did; each member now has their own slot. And the archiving policy you set per collection now applies on every path: previously only the nightly scheduler read it, and the manual sweep could archive never-accessed documents at half the intended age.

Fix

Plan changes no longer create a second subscription or double-charge; downgrading to Free cancels the previous billing. Member permissions closed (passwords, tokens, gaps). Provisioning and sync workers stable. Digest and health cards now count everything awaiting a decision.

Feature

Impact home with counted value for the period: answered searches, delivered contexts, and recorded knowledge, by function. Daily e-mail digest for the workspace owner, sent in the tenant's timezone, with click tracking.

Spanish landing, /research and /vs/hindsight

Feature · August 15, 2026

Full public landing under /es. 400-question benchmark study on /research and competitor comparison on /vs/hindsight, with the benchmark synced to 96.58% R@10.

Faster dashboard on hot routes

Improved · August 15, 2026

60s cache on engine reads and single-flight fetch dedup (30s TTL) across the dashboard's five most-visited GET routes.

Security and test-suite hygiene

Fix · August 15, 2026

Security review fixes (critical route auth and adjacent ones). The test suite now runs against the test database only — production stays untouched. Hydration error in the verify widget resolved.

Feature

Side navigation by Account / Workspace / Billing / Developer. Hub at /settings with shortcuts. Mobile uses a section select. Gone is the confusing chip bar at the top.

Ask memory at /knowledge/ask + confirm origin

Feature

Friendly URL (legacy ?tab=ask and /knowledge/chat redirect). Deep-link /knowledge?veracity=imported to review imports. On doc detail: “Confirm by team” (veracity=stated) closes the trust-score loop.

Multi-tenant retrieval ranking (contextPolicy)

Feature

At /settings/retrieval: prefer/demote collections, paths and content types, plus purposeBoosts. No brand hardcodes in the engine. Sticky save while scrolling.

Landing + satellites in Hallmark craft

Feature

Split hero, comparison table, pricing matrix, workbench FAQ. /use-cases and /vs redesigned: directory rows, row comparison (no twin red/green cards), feature matrix first on vs pages.

Engine: PATCH veracity + listing with origin

Infra

GET /documents exposes veracity; PATCH /documents/:id/veracity for human confirmation. SaaS browse filters by origin. Rank harness via tenants.settings.contextPolicy (tenant-agnostic).

Feature

Hero with value stats (decisions, embedding coverage). TeamThisWeekCard. Analytics with agent value. ROI widget with honest onboarding. Empty states on widgets.

/roi page + /vs + Team agents

Feature

Visual value comparison. /vs index. Settings/users with per-agent contributions. Graduated onboarding progress.

CI green + fetch hardening

Fix

Stable CI on engine and SaaS. safeJsonFetch on critical paths. i18n and a11y (aria-labels). Loading.tsx on protected routes. Published the first LoCoMo pipeline study, later re-baselined by the canonical table.

Feature

Verification cron on day 7. /en/pricing with SubscribeCard, FAQ and calculator in parity with PT.

Feature

Connectors for Notion, Slack, Drive, GitHub and Linear. /onboarding/connect-agent wizard. Verify Connection widget on the dashboard.

Engine billing wired + email activation

Feature

recordUsage on critical paths. /auth/activate with full state machine. Paid plans require verified email.

Health monitor + Sentry scaffolding

Infra

Health timer with Telegram alerts. Optional Sentry (engine + SaaS) with tenant tags.

Polished internal pages + partial EN

Feature

Knowledge with filters; canonical with diff; audit-logs with export; API keys with scopes. Main /en/* pages translated.

Feature

Checkout, credit packs, subscription UI. Landing, pricing, comparison, use-cases, vs. Multi-tenant Postgres RLS (4 layers) on the engine.

Feature · July 16, 2026

With indexed memory, login and the main nav open Ask (/knowledge/ask). Dashboard, Ctrl+K, and the first-source tour push cited questions. Browsing memory stays secondary.

Weekly trust sample

Feature · July 16, 2026

The trust card asks for a small sample (up to ~12 decisions and facts), with a clear count. Team confirm updates the weekly sample.

Sync: what landed in memory

Improved · July 16, 2026

After OAuth and sync, status shows how many documents landed per source, with a shortcut to Ask. The connect toast notes that the first sync is starting.

Refined look, real hero, locked design system

Improved · July 18, 2026

Light by default (dark one toggle away). Landing shows a real dashboard screenshot, without animated mock or fake browser chrome. Clearer Ask and memory subnav. Typography, mono metadata, brand green as signal. Rules locked in design.md for the next screens.

Feature · July 14, 2026

Trust / Economy / Performance cards on the dashboard. Knowledge gaps with dismiss/fill. Onboarding goal “1 source + 1 cited answer”. Memory-first positioning (not “virtual company”).

Veracity (origin) on search, list, and Ask

Feature · July 13, 2026

Origin badges: team-confirmed, integration, agent, suggested. Filter by veracity. Ask sources show origin and score. Trust KPIs in analytics.

RBAC: team-lead vs personal view

Feature · July 14, 2026

Dashboard and agent presence differ for team-lead vs solo/founder. Real team activity on the Valor Digital workspace (dogfood).

Knowledge assistant (Ask) GA in product

Feature · July 9, 2026

Memory chat, dashboard widgets, i18n. MCP tokens with multi-user UX and Hermes snippets. Unified engine gateway (engineRawFetch).