Changelog | Mandaire dev

Date	Component	What shipped	Who it affects
2026-06-25	API	MCP response envelope compaction (v1.13). Static blobs moved to a manifest referenced by the response rather than duplicated per-response. Smaller payload per call; field structure unchanged. Callers should update parsers to handle the `manifest` reference field if reading static content directly.	All developers; renderer LLMs
2026-06-25	UX	Query responses now cap at 15 results with a drill-in pointer, down from 30. The first call returns the most relevant results; callers can request more via the drill-in pattern. Applies to message-route queries. No change to structured from_kind queries.	Developers consuming message-route responses; renderer LLMs
2026-06-25	ER	File entity resolver live: people mentioned in documents and files are now resolved against the entity graph. If a name in a file matches a known entity in the user's history, the connection is made. Expands get_person coverage to document-referenced contacts.	.app operators; developers using get_person
2026-06-23	Graph	Mediated relationship propagation fixed. Friend-of-spouse, colleague's partner, and similar second-degree relationships were not propagating through the entity graph. Now fixed: the graph reflects the full inferred social graph, not just first-degree edges. Affects get_person responses for entities known through a mediating contact.	.app operators; developers using get_person
2026-06-13	Docs	Published "Three things to build when your AI knows the person." Three concrete implementation patterns with actual MCP calls: pre-meeting context enrichment (`from_kind="pre_engagement"`), per-recipient disclosure filtering (how `disclosure_applied` and `absent_knowledge_caveats` work for multi-viewer products), and the read-not-store query pattern with context discarded after session. At mandaire.com/blog/build-patterns/.	Developers building on Mandaire
2026-06-13	Docs	Developer FAQ live at mandaire.dev/faq/. Covers the 12 most common integration questions: auth setup, empty query results and when to check `disclosure_applied` vs `absent_knowledge_caveats`, writeback slots, confidence values on INSERTs, and when `purpose=` is required. Sourced from production behavior; updated as patterns emerge.	Developers integrating Mandaire
2026-06-13	Docs	External brand term locked: "personal knowledge graph" is the public-facing term for the core product layer. "Governed personal graph" is internal-only (ADRs, technical docs, cross-layer dispatches). Updated all four surfaces. Developers should use "personal knowledge graph" in renderer-facing copy.	Developers; all surfaces
2026-06-13	MCP docs	MCP docs page restored to mandaire.dev/for-developers/. The page covers the full verb shape, five tools, response envelope fields, nil-result writeback slot, INSERT required fields, anti-patterns, and quick reference. Previously linked off-domain; now local.	Developers connecting renderer LLMs
2026-06-13	ER	Entity resolution accuracy reviewed for the "Who is Jordan?" documentation post. Three-signal ER (shared identifiers, co-occurrence patterns, communication history), personal entity ranking, and contamination detection cadence all verified against production code paths. A fourth claim, personal knowledge graph terminology, updated to match the locked external term.	.app operators; developers reading docs
2026-06-10	Privacy	Privacy request page live at mandaire.com/privacy/request. Handles data access, deletion, and correction requests under CCPA and GDPR timelines. Operators who have built on Mandaire can point users here.	All users; .dev operators
2026-06-10	Docs	Published "Your app does not need to store your users' history." A developer post on the read-not-store architecture: MCP query at runtime, discard after session, disclosure engine runs upstream before your AI sees the result. No user data stored on your end, no compliance surface for context you never held. Links directly to the MCP docs.	Developers evaluating Mandaire
2026-06-07	UX	Connector approval labels now distinguish between request types. A call for private briefings shows a different label than a read-only query call. Previously identical labels made it hard to tell what kind of access was being requested. The data accessed has not changed; the transparency about what is being requested has.	.app users; .dev operators
2026-06-03	API	Every query response now includes a `retrieval_status` field. Values: `found` (information returned), `not_indexed` (source not yet ingested or indexed), `withheld` (policy blocked the result). Previously, an empty result was ambiguous between these three states. Callers should branch on this field rather than testing for empty content.	All developers; renderer LLMs consuming MCP responses
2026-06-03	API	Recall queries now auto-widen their time window when the query implies a range that would miss relevant context. The response includes a caveat field when widening occurred. Callers that set an explicit time range will not be widened; auto-widening only applies when the window is inferred from the query.	Developers using recall-type queries
2026-06-02	API	Four new `from_kind` values: `decision_brief` (full history behind a choice), `commitment_risk` (what was promised and whether it is on track), `since_last_seen` (catch-up summary before a meeting), `precedent` (how a situation has played out in history before). Each returns a structured response against the same envelope shape as other from_kinds. All are read-only; no write-back required.	Developers building pre-meeting or commitment-tracking features
2026-06-05	ER	Entity disambiguation improved for high-confidence cases. If you query "who is Alex?" and one entity named Alex accounts for most of the co-occurrence signal in the user's history, Mandaire returns that entity without asking which one you mean. Disambiguation prompt still appears when the signal is genuinely split between two candidates.	.app operators; renderer LLMs
2026-06-09	Positioning	WWDC Extensions Framework confirmed: the AI brain is a swappable commodity. Updated all four surfaces (.com/.app/.dev/.org) to the "bring your own brain" frame. Mandaire is the context layer any AI brain plugs into via MCP. The corpus moat argument is the primary positioning claim going forward.	All developers evaluating Mandaire
2026-05-30	Docs	Added "Get started" to the navigation on all .dev pages. Links to the setup guide. Previously the setup guide existed but was not linked from anywhere; developers landing from search had no signposted path to it.	Developers evaluating Mandaire
2026-05-28	API	Nil-result SELECT envelopes now include a `writeback_slot_id`. When a query returns empty, your AI can write back what it learns in the conversation using an INSERT with that slot, closing the loop from "not in the graph yet" to "now it is."	.app operators
2026-05-28	Positioning	Repositioned across all four sites as a personal knowledge graph. Named the three structural failures LLM memory cannot fix: deduplication, recency weighting, per-recipient disclosure. Read-only architecture and write-back compounding now explicit in the copy.	All sites (visitors)
2026-05-27	Disclosure	Per-recipient filtering enforced for health and employment topics. All other topics in shadow evaluation (logged, not enforced).	All .app users
2026-05-19	Core	OAuth 2.1 + PKCE. First authenticated API endpoint live for private beta.	.dev beta users

What shipped. When it shipped. Who it affects.