What's New
Stay up to date with the latest features and improvements to Hindsight Cloud.
October 13, 2026
Hindsight 1.0
After a year in production, Hindsight's architecture and API are stable, and from here the API is a promise: what you build on Hindsight Cloud keeps working as Hindsight grows. Nothing changes in your code, SDKs or MCP setup.
Hindsight Memlake
Hindsight Memlake is a new storage engine for agent memory, written in Rust, that keeps memory in object storage: 100 million memories at under 100 ms p99, with every write durable before it's acknowledged. Most of Hindsight Cloud already runs on Memlake, and Vectorize operates it for you, so there is nothing to turn on. Read the launch post for how it works, or see Hindsight Memlake for running it in your own cloud or data center with Hindsight Enterprise.
October 2, 2026
Reflect Can Read Document Metadata
Reflect now sees each document's own metadata. If you retain documents with metadata that says where they came from, for example {"doc_type": "kb"} for published articles and {"doc_type": "ticket"} for support conversations, a bank's mission can tell reflect how to rank them: prefer the handbook over the chatter, unless something newer and authoritative overrides it.
Members Can Connect AI Clients
Organization owners can now let members approve MCP connections from AI clients such as Claude, ChatGPT and Cursor. Before, only owners and admins could. Turn it on under Settings → Connected AI clients (MCP).
Connections now follow access, too: when someone leaves the organization, or loses the role that let them connect, their MCP connections are revoked. And signing in with SSO during an MCP connection now returns you to the approval screen instead of the dashboard.
Tag Filters for the Knowledge Base and Entities
The knowledge base tree and search, and the entity list, detail and graph endpoints, now accept the same tag filter as recall (tags with tags_match). Use it to show only the pages and entities for one user, team or topic.
Learn more about the knowledge base →
September 30, 2026
Bank Aliases
A bank can now answer to more than one name. Add an alias to a bank and requests to either id reach the same memories. Promote an alias to make it the name the bank is shown as everywhere.
Aliases make renaming painless: point new clients at the new name while existing ones keep working on the old one, with no downtime and no data to move. API keys scoped to a bank work with its aliases too. Owners and admins manage aliases on the bank's Configuration tab.
Better Evidence Behind Reflect Answers
The Based On panel for a reflect answer now shows the observations it cited, with their full text, and the images and files attached to its evidence, so you can see exactly what an answer drew on.
Find Documents Faster
- A time filter on the Documents list, with presets or a custom range, and a tag filter
- Document metadata shown on each memory's details
- An advanced search dialog for knowledge base pages
- Causal links in bank stats, and pause, motion and SVG export for the constellation graph
The memories and documents APIs also filter and sort by time on the server.
ChatGPT Connector
ChatGPT's custom connectors can now link to Hindsight Cloud over MCP. Add Hindsight as a connector in ChatGPT, approve it once, and ChatGPT can retain and recall memories in your banks.
September 19, 2026
Images and Files in Your Memories
Images and files attached to retained content now display in place in the memory, document and chunk views, instead of placeholder text. Recall results show the attachments behind each fact, so an agent can get the screenshot or diagram along with the text.
See Exactly What Hindsight Sends
The bank Configuration tab can now preview the exact prompts that retain, consolidation and reflect will send, with your bank's mission, directives and settings applied. You can also run a free, dry-run extraction test on sample text to see what facts and entities retain would extract, before you change anything. Owners and admins can save edits; members see the configuration read-only.
More Control Over Banks and Recall
- A per-bank switch for keyword search in recall
- A default refresh setting for new knowledge pages in each bank
- Include knowledge base when exporting a bank
- Recall results show entity and tag chips, timestamps and more precise scores, with help text on every recall and reflect option
- Cancel an in-flight operation
- The current bank is pinned at the top of the bank selector
Data Privacy (GDPR)
Organizations on Enterprise plans can now erase a data subject's memories, run DSAR exports, set retention rules, review a Privacy Events ledger, and redact EU personal data at retain time.
Learn more about data privacy →
September 8, 2026
VAT IDs on Invoices
Add your VAT ID to your billing details and it's printed on your invoices. Each invoice keeps the details it was issued with, and saving your billing details also fills them in on invoices issued without them.
Lower Cost to List Mental Models
Listing mental models, through the API or the MCP list_mental_models tool, now returns each model's details without its content. A list no longer reads every model's full text, so it no longer costs a Retrieve Model read for each one. Fetch a model directly to read its content.
Learn more about mental models →
Enterprise Support
Organizations on an enterprise support plan now have a ticket portal under Support, email intake at support@vectorize.io, and contracted response times.
Learn more about Enterprise Support →
September 1, 2026
Simpler Google Workspace SSO
You can now set up SAML single sign-on with Google Workspace by pasting the three values Google shows you (SSO URL, Entity ID and certificate), without needing a metadata URL. Organizations with SSO also block password sign-in and password resets for addresses on their SSO domains, so everyone signs in through your identity provider. SSO is available on Enterprise plans.
August 27, 2026
Knowledge Base
Your memories now have a place to live as readable, organized documents. The new Knowledge base view lets you build folders and pages inside a bank — a handbook that writes itself from what the bank already knows.
A page isn't something you type. You give it a source query — the question that rebuilds the page — and Hindsight synthesizes the answer from the bank's memories. Ask "What are our refund and cancellation policies?" and you get a page. As the bank learns more, the page follows.
Because a page is defined by its question rather than its text, it never goes out of date the way a hand-written document does. Change the question and the page rebuilds; leave it alone and it keeps pace with your memories on its own.
What you'll see:
- A Knowledge base entry in the bank sidebar, alongside your memories and mental models
- New folder and New page actions, with pages nested wherever you want them
- A source query field on every page — "the question that rebuilds this page" — plus Tags for cross-linking; a
type:...tag sets the page's type - Generating... while a new page is being synthesized, then the finished page with How this page is derived and Backed by N memories so you can see what it drew on
- A Stale or In sync badge on every page, telling you whether new memories have landed in that page's scope since it last read them. Pages that refresh themselves clear their own badge; the rest tell you when a refresh is worth running
- Search across every page in the bank
- Editing a page's name, question or tags — changing the question rebuilds its content
- Deleting a folder removes everything inside it, after a confirmation naming what will go
Organization members with read-only access can browse the knowledge base but not change it, the same as everywhere else in Hindsight.
The knowledge base is also fully available over the API, under /v1/default/banks/{bank_id}/knowledge-base/ — read the tree, create folders and pages, search, and export the whole base. See the API reference for details.
What it costs: a page is a mental model underneath, and it bills on the meters that already cover them — no new charge type. Building or rebuilding a page is a single Refresh Model call ($0.05); reading one is a Retrieve Model read ($0.25 per million output tokens), the same as fetching the model directly.
Learn more about the knowledge base → | See Refresh Model pricing →
July 7, 2026
Scheduled Mental Model Refresh
Mental models can now refresh on a schedule you control. Each model has a refresh trigger with three modes:
- On new memories — refreshes automatically when new memories change the model's scope (the existing default behavior)
- On a schedule (cron) — refreshes on a fixed UTC cron schedule you define, such as
0 3 * * *for daily at 03:00 UTC - Manual only — refreshes only when you trigger it by hand or via the API
Scheduled mode is built to be cost-safe: a scheduled refresh only runs when new memories have actually changed the model's scope since the last refresh. If nothing changed, the scheduled run is skipped and you aren't billed for it. This makes cron scheduling a predictable-cadence alternative to on-new-memories mode — you choose when updates land instead of having them follow consolidation activity — without paying for no-op refreshes.
What you'll see:
- A refresh trigger selector when creating a mental model, and in the model's detail view for existing models
- A live cron preview as you type a schedule — human-readable description plus the next upcoming runs in both UTC and your local time
- A Next refresh indicator on each model showing when it will update next ("on new memories", a scheduled time, or "on demand")
- A Scheduled badge on models that refresh on a cron schedule
Learn more about mental models →
Export a Bank's Documents
You can now export a bank's documents — along with the memories extracted from them — as a single downloadable .zip archive. Each export bundles the document text, extracted facts, entity names, causal links, and chunks, with an option to also include consolidated observations.
Use it to back up a bank or keep an offline copy of everything a bank has learned.
What you'll see:
- An Export action in a bank's Documents view
- A dialog with an option to include consolidated observations
- A downloaded
your-bank-documents.ziparchive
July 6, 2026
Updated Pricing Model
We've updated Hindsight Cloud's pricing to better align what you pay with what each operation actually costs to run — and to make Hindsight Cloud more cost-effective overall. Retain — the most-used operation in nearly every workload — is now 33% cheaper. The new rates take effect for all usage from July 6, 2026 onward; activity before that date was billed at the legacy rates and is unaffected.
What changed:
- Retain is now $10.00 per million input tokens — down from $15.00, a 33% reduction. For most workloads this is the single largest line on your bill.
- Reflect and Refresh Model are now charged a flat $0.05 per call (previously $3.00 per million output tokens). These operations run an internal agent loop whose cost is closely tied to the number of calls rather than to the size of the visible output, so per-call pricing more accurately reflects what we pay to deliver them.
- Storage is a new line: $0.25 per million tokens stored per month. Memories are free for their first 30 days, and only the portion older than that counts toward storage. Your memories are never deleted.
- Recall, Retrieve Model, and Iris rates are unchanged.
What you'll see:
- The Billing page now shows the effective rate for each operation in its native unit (per call, per million tokens, or per million tokens per month for storage).
- Per-operation cost cards show what you were actually billed — not a recomputation at current rates. Charges from before July 6 stay at the legacy rates; activity from July 6 onward uses the new rates. Sums always match your actual balance movement.
- Dollar amounts display at two decimal places; hover any amount to see full sub-cent precision.
See full pricing → | Read about storage billing →
July 1, 2026
Refreshed Look and Feel
Hindsight Cloud has a refreshed visual design across every page — cleaner typography, tighter spacing, more consistent components, and a smoother experience top to bottom. No workflow changes; everything you already do works the same way.
June 16, 2026
Memory Defense
Memory Defense screens retain calls before content reaches storage, so secrets, prompt injections, and tampering attempts never enter your memory bank. It's enabled per bank — a bank admin opts in and chooses which protections to run.
Memory Defense protects against the three highest-impact attack families on agent memory:
- 🔑 Secret exfiltration — credentials, API tokens, and access codes that leak into chat or tool output get caught before they're stored
- 💉 Prompt injection — attacker-controlled content embedded in scraped pages, tool outputs, or planted memories that would otherwise hijack a future session (OWASP ASI06 Memory Poisoning)
- 🧪 Integrity tampering — writes under trusted tags or flooding attempts that distort the agent's view of reality
Bank admins configure the policy from Bank Settings → Memory Defense, choosing the detectors that match their threat model and whether each one redacts content or blocks the retain entirely.
Tier availability: regex-based credential redaction (44 common API key patterns) is available to everyone. The full detector suite — extended provider catalog (170+ additional credential patterns), base64-encoded secret detection, conversational secret detection via LLM, prompt injection blocking, size anomaly blocking, and protected tag namespaces — is Cloud Enterprise. See the feature matrix in the overview for the full breakdown.
Learn more about Memory Defense →
May 15, 2026
Alipay at Checkout
You can now pay for credits with Alipay (支付宝) in addition to credit and debit cards. Alipay is offered at Stripe Checkout to customers in mainland China — the exact payment options shown depend on your locale and merchant configuration.
- Single-use authorization — Alipay payments are authorized one charge at a time, so they aren't saved to your account
- Auto-recharge is card-only — to enable auto-recharge you'll still need to add a card; you can do this from the Payment Methods card after an Alipay purchase
- Existing card flow is unchanged — Visa, Mastercard, American Express, and other major networks continue to work exactly as before
Learn more about accepted payment methods →
Multilingual Control Plane
The Hindsight Cloud UI is now fully internationalized. Use the language picker in the top navigation bar to switch between 8 supported locales: English, Spanish, French, German, Portuguese, Japanese, Korean, and Chinese (Simplified).
- Every page and dialog is translated — dashboard, billing, settings, team, API keys, usage, support chat, admin, and all bank views
- Per-browser preference — your selected language persists across sessions
- Documentation in Simplified Chinese — this documentation site is also available in 简体中文; switch via the language picker in the docs header
Learn more about language preferences →
May 6, 2026
Constellation Graph View for Memories and Entities
The Memories Data view has a new Constellation visualization that lays out memories and their relationships as an interactive graph, and the Entities view gains a new Relations tab using the same engine to explore entity-to-entity connections.
- Interactive layout — zoom, pan, and click any node to inspect the underlying memory or entity
- Color-coded fact types so World Facts, Experiences, and Observations are easy to tell apart at a glance
- Legend overlay explains each node and edge type
Richer Memories View
The Memories Data view picks up several quality-of-life improvements.
- Tag filtering — filter memories by one or more tags directly from the view
- Server-side text search — search across all memories in a bank, not just the current page, via a new
qquery parameter the UI uses under the hood - Memories-ingested chart with a period switcher (24h / 7d / 30d) on the bank stats page so you can see ingestion trends over time
- Failed-consolidation drilldown — click a failure on the Consolidation card to see exactly which inputs failed and why
- Empty-state CTA — banks with no memories now surface an Add Document button right in the data view
Mental Models View, Fully Refreshed
The Mental Models view has been rebuilt to surface staleness and observation history.
- Staleness indicators show when a model is older than the most recent memories that informed it, with a one-click refresh
- Recall controls let you preview what a model returns and tune its source query without leaving the view
- Near-fullscreen detail modal with CompactMarkdown previews on each card and an Observation History panel inside the memory detail to trace how a model evolved
- Directive detail modal in the Think view for inspecting individual directives in context
Operations Dialog Upgrades
The async Operations dialog gets a set of long-requested controls.
- Retry failed file-extraction operations directly from the row, no need to re-upload
- Cancel pending operations, both from individual rows and from a new dialog-level Cancel action
- Task-type filter narrows the list to a single operation kind (consolidation, file extraction, mental model refresh, etc.)
- Cancelled and Processing status pills for clearer at-a-glance state
- Load raw and metadata JSON for any operation when you need to debug
Export and Import Bank Templates
The bank Actions menu now includes Export Template and Import Template so you can save a bank's mission, directives, traits, and structural settings as a portable file and apply them to a new bank.
- One-click export produces a JSON template you can check into version control
- Import on creation lets you spin up a new bank pre-configured to match an existing one
- Recover Consolidation action — also new in the Actions menu — re-runs consolidation for a bank when observations have drifted out of sync with the underlying facts
Refreshed Look and Feel
A round of UI primitive updates (cards, tables, status pills, dialogs) gives the entire control plane a tighter, more consistent visual style. Every view that uses these shared components inherits the new look automatically.
March 29, 2026
Multi-Organization Support
You can now create multiple organizations from the organization switcher in the top navigation bar — useful for separating production and development environments or managing distinct projects.
- Instant switching — click the org name in the header to switch between your organizations at any time
- One free credit grant per user — welcome credits are tied to your user account, not each organization, so they are only granted once regardless of how many orgs you create
Learn more about creating organizations →
Credit Transfers Between Organizations
Owners can now transfer credits between organizations they own directly from the Billing page — no need to contact support.
- Instant transfers — credits move atomically; both organizations' balances update immediately
- Full transaction history — the source org records a
transfer_outentry and the destination org records atransfer_inentry, visible on each org's billing page - Confirmation step — a confirmation dialog shows the amount and destination before any funds move
- Owner-only — only users who own both the source and destination organization can initiate a transfer
Learn more about credit transfers →
March 23, 2026
MCP OAuth 2.0 with Dynamic Clients
MCP connections can now use OAuth 2.0 with dynamic client registration instead of relying only on static API keys.
- Native OAuth flow — Hindsight now handles MCP OAuth approvals directly in the control plane
- Dynamic client registration — compatible MCP clients can register themselves automatically during setup
- Browser session reuse — if you're already signed in to Hindsight, authorizing a new MCP client is faster
- Explicit approval step — every MCP authorization now clearly shows what access you are granting
- Organization-aware controls — only org admins and owners can approve MCP access
March 10, 2026
Programmatic API Key Management
API keys with the Key Creator capability can now programmatically create, list, and revoke bank-scoped child keys via the API — no admin or UI access required.
- Create child keys — provision short-lived, least-privilege keys scoped to specific banks
- Bank scope enforcement — child keys can only access banks within the parent key's scope
- Expiration constraints — child keys cannot outlive their parent
- Cascade revocation — revoking a parent key automatically revokes all of its children
- Immutable children — programmatically created keys cannot have their bank scope edited; revoke and recreate instead
- Audit trail — all key creation, revocation, and scope changes are logged with actor identity
Learn more about programmatic key creation →
March 9, 2026
Document File Upload
You can now upload files directly to your memory banks. The Add Document dialog supports PDFs, Word documents, PowerPoint presentations, Excel spreadsheets, images, and plain text files.
Choose between two extraction methods when uploading:
- Standard — Free text extraction using Markitdown, ideal for most text-based documents
- Enhanced (Iris) — AI-powered extraction with an additional per-token charge, best for images, scanned PDFs, and complex layouts
After uploading, track progress in the Document Operations panel — each operation shows its status (Pending, Completed, or Failed) so you always know where things stand.
Learn more about documents → | View Iris pricing →
March 3, 2026
Bank-Scoped API Keys
API keys can now be restricted to specific memory banks, giving you fine-grained control over which banks each key can access. If a restricted key attempts to access an unauthorized bank, the request is rejected with a 403 error.
- Create scoped keys — select which banks a key can access during creation
- Edit restrictions — update bank access on existing keys at any time
- Enforced everywhere — restrictions apply across all HTTP endpoints and MCP tools
Learn more about bank-scoped keys →
February 13, 2026
MCP Support
Hindsight now supports the Model Context Protocol (MCP), letting you connect your memory banks to any MCP-compatible AI client including Claude Code, Claude Desktop, Cursor, Windsurf, and VS Code.
- Single-bank mode for dedicated agent memory scoped to one bank
- Multi-bank mode for agents that need to work across multiple banks
- 30 tools available including retain, recall, reflect, mental models, directives, memory browsing, documents, operations, and bank management
- One-click configuration from the Connect page in Hindsight Cloud