What has shipped. Mandaire is in active development. This page tracks meaningful changes to the product: new capabilities, architectural shifts, and surface updates.
The claim-store stats on mandaire.app/proof and mandaire.org had drifted from the live measurement: 22,226 live claims was actually 22,146, and the corroboration rate cited against a specific corpus artifact had moved from 56% to 54.4%. Both pages now reflect the 2026-08-07 measurement.
mandaire.com and mandaire.app now lead with "Mandaire reasons over your context instead of just storing and retrieving it" — in the .app hero subtitle and in the social-card description on both sites. A direct response to competing products that frame themselves around storing and retrieving your context rather than reasoning over it.
LLM memory is a chat history. A knowledge graph is something else: resolved identities, calibrated inferences, deterministic disclosure rules. Three structural gaps no larger context window closes. Read it.
Claude Code, Cursor, Windsurf, and Copilot all reset context every session. The decisions made yesterday, the constraints discovered, the approaches ruled out: gone. What a persistent knowledge layer changes about that. Read it.
The FAQ and homepage previously said "export at any time" without making clear that export works today. Updated to "Export runs today": the full graph, inferences, and raw corpus can be requested via the privacy request page. The portability claim was architecturally true; the copy was not specific enough to be trusted at face value. An external cold review (Perplexity, June 2026) flagged this as the gap between "no lock-in" as a claim versus a demonstrated fact. The export mechanism has been live since the product launched.
Two FAQ pages are now live. mandaire.com/faq covers the common questions before people decide to request access: what happens to their data, how Mandaire differs from AI memory, which AI systems work with it, and what "private beta" means in practice. mandaire.dev/faq covers developer integration questions: auth, empty query results, writeback slots, confidence values on writes, and when the purpose= parameter is required.
The daily brief was returning an empty response on the iOS Siri Shortcuts path since late May. The web client was unaffected. Root cause: when the rendered_text field was added to the response contract, it was added to the main compose path but not the fast-slice path that runs within the Shortcuts latency ceiling. The fast-slice returned a valid envelope with an empty field. The Shortcuts client read the empty field and said it had nothing.
Both code paths now produce rendered output. The brief is available on iOS again. The failure is named in the failure admissions record.
The install page at mandaire.com/install now includes setup instructions for ChatGPT alongside the existing Claude.ai steps. Select your AI from the tab at the top of the setup section; the steps adjust accordingly. The canonical MCP endpoint is mcp.mandaire.com for Claude and mcp.mandaire.com/mcp for ChatGPT. The old alias (mcp.mandaire.app) that appeared in earlier copy has been corrected across the site.
Architecture copy on mandaire.com/architecture and a draft post stated that raw source data is discarded after indexing. That is the destination architecture, not the current behavior. Today, Mandaire retains the raw corpus locally on your server alongside the knowledge graph, to power live retrieval. Both stay within your infrastructure; neither leaves to a third party. The goal architecture eliminates local raw retention as well, with on-device key derivation so only you hold the decryption key. The copy now describes both: current behavior and where we are building. The key-derivation framing has been consistent since the architecture page launched; this correction brings the retention framing into alignment with the same hedging standard.
The external product term for the graph architecture is "personal knowledge graph." An earlier update (June 11) used "governed personal graph" as the primary description. That term works as internal engineering shorthand but reads as compliance jargon on public-facing copy. The governance layer is real and architecturally significant; the word "governed" is not the right carrier of that significance for a consumer product. "Personal knowledge graph" is precise and understandable. The copy across all four surfaces now uses this term where the architecture is referenced.
A developer post with actual MCP calls showing three concrete patterns: pre-meeting context enrichment via from_kind="pre_engagement", per-recipient disclosure filtering that surfaces how disclosure_applied and absent_knowledge_caveats work for multi-viewer products, and the read-not-store query pattern. The first post on the blog to show the wire in practice rather than describe the architecture.
The most common question before someone requests access is what actually happens when you connect Mandaire to your Gmail or calendar. The post explains what OAuth means in practice, what Mandaire processes at ingest, what it stores (the knowledge graph, not your messages), and what your revocation and deletion rights are. Written for people who want to understand before they decide.
The product is now described as a "governed personal graph" across all four surfaces. Earlier copy in several sections used "personal knowledge graph" as the primary description. The change is deliberate: "governed" is the architectural differentiator. The graph has deterministic disclosure rules enforced before any AI query. A knowledge graph without governance produces structurally different outcomes. The copy now reflects that.
Superseded June 13: "governed personal graph" is now internal-only. External copy uses "personal knowledge graph."
You can now submit data access, deletion, and correction requests directly at mandaire.com/privacy/request. Requests are responded to within 45 days under CCPA and 30 days under GDPR. The link is also in the site footer.
The mandaire.app product page now describes what the knowledge graph actually contains: every contact understood across every source, who they are, what they care about, what your relationship with them is, what you should share with them, and which claims from your history are reliable enough to believe. Deterministic disclosure rules. Switch AI providers without losing any of it.
A developer-targeted post on the read-not-store architecture. Your AI renderer connects to the user's context store via MCP, reads what it needs, and discards it when the session ends. No user data stored on your end, no deletion workflows, no breach liability for context you never held. The disclosure engine runs upstream: the result arrives pre-filtered.
WWDC confirmed the argument we have been making: Apple spent a billion dollars a year renting Gemini's reasoning because building intelligence is not the durable problem. What a swappable brain cannot take with it is your history. We wrote about what that means for who holds the durable advantage in personal AI.
We updated the positioning across all four surfaces to reflect what Apple confirmed at WWDC. The new frame: Apple built the interface for choosing your AI's reasoning engine. Mandaire holds the context that brain needs, on a server you own. The brain is swappable. Your 20-year correspondence history is not.
The hero on mandaire.com was also updated to lead with the disclosure moat. The draft you send your investor should not be the draft you send your spouse. Mandaire is the product that makes this structurally enforced, not just hoped for.
When you connect Mandaire to Claude or ChatGPT, the tool approval prompt now labels each access type separately. A query for private briefings shows "Mandaire primary data query." A read-only call shows a distinct label. The prompts were previously both labeled the same way, which made it hard to tell what kind of access you were approving.
The data it reads has not changed. The transparency around what is being requested has.
Asking "who is Alex?" now returns your actual close contact first. If there is only one person in your history who accounts for most of the signal on that name, Mandaire no longer surfaces a disambiguation question. Asking about someone returns the person you mean.
Asking "what did we discuss about the deal?" now works across the full conversation arc, not just the most recent context. If you have months of email on a topic with someone, the answer draws from all of it.
Four new query types are available via MCP. Ask for a decision brief when you need the full history on a choice you are facing. Ask for a commitment risk when you want to know what you have promised someone and whether you are on track. Ask for a since-last-seen summary when you need to catch up before seeing someone. Ask for precedent when you want to know how a situation has played out before in your history. Each is a deterministic query against your knowledge graph; no conversation warmup required.
Every query response now includes a retrieval_status field that tells you exactly why you got the result you got: the information was found, the information is not yet indexed, or the policy withheld it. You no longer have to guess when Mandaire says nothing whether it found nothing or chose not to say.
Recall queries also handle uncertain date ranges better. When your question implies a time window that would miss relevant context, the query automatically widens and flags that it did. The caveat is in the response rather than something you have to notice by checking manually.
The architecture page previously described the zero-knowledge encryption design goal in a way that could be read as a current-state claim. The corrected copy now states explicitly: "We have not yet completed the zero-knowledge architecture that achieves this." What is in place today and what is being built toward are described separately.
If you are evaluating Mandaire for sensitive use, the architecture page now gives an accurate read on where the encryption model stands.
We rewrote the positioning across all four sites. The prior framing described Mandaire as a "memory layer" (accurate but incomplete). The new copy names the three structural failures of LLM memory: deduplication (it accumulates references without resolving them), recency (it weights recent conversations, not relationship depth), and disclosure (it cannot answer the same question differently depending on who is asking).
These are not scale problems. No amount of context window expansion fixes them. The copy now makes that explicit.
Additional update: the architecture copy now reflects that Mandaire is read-only from underlying sources. Your AI reads from Mandaire via MCP; if it needs to act (draft an email, update a calendar), it does so through a separate connector. Mandaire informs. It does not act.
The first external viewer account is live. A viewer can connect to a Mandaire instance via any MCP-compatible AI (ChatGPT, Claude, or others) and read briefings the account holder has chosen to share. No separate login, no new app: the viewer connects through the AI they already use.
This is the first time someone other than the account holder has seen Mandaire briefings. The disclosure engine governs what surfaces per viewer; the account holder sets the policy.
We clarified the product architecture across all four surfaces. Mandaire is the memory layer. You bring your own AI. Claude, ChatGPT, Gemini, or any MCP-compatible model reads from the same private store through a single connection. The AI does not change; what changes is that it now has context it never had before.
This framing better reflects how the product actually works. The previous copy described features in isolation; the updated copy describes the architecture as a whole.
We rewrote the mandaire.app homepage for a mainstream audience. The prior version described what the product is technically. The new version leads with what it does for people: handle the digital so you can show up for the human parts.
The page now leads with the network value (every person you add makes the whole thing more useful for everyone), rather than the architecture. Technical readers still find what they need; the primary story is now accessible to anyone.