INITE Brain

What this server is and what it can do, before you connect. Reads nothing and stores nothing.

Dry-run preflight against fn::resolve_fact. Answers "if I were to record this fact right now, what would the resolver decide?" without writing to the database. wouldOutcome ∈ {INSERTED, SUPERSEDED, COMPETING, REJECTED}; reasoning explains which rule fired (semantics class, score gap vs margin, cosine threshold, etc); opposingFacts lists the same-predicate priors the resolver would have weighed against. Useful where the cost of a contested write is high (e.g. agent loops that pay an ingest credit). Fidelity: this preflight uses the SEED source-trust table and authority 0 — the live resolver scores opponents with the learned domain-scoped rates (migration 0045) and declared source authority (migration 0046), so predictions can differ for sources with conflict history or a registered authLevel.

Returns the topic communities a given entity is a member of. Places an entity in its broader context — which clusters / themes it participates in — and leads from one entity to the wider neighbourhood the cluster summarises.

Get entities connected to the given one via the knowledge graph.

Returns facts in COMPETING status — those the conflict resolver couldn't auto-supersede because two same-predicate bitemporal facts overlap in valid-time and are too cosine-close within margin. Grouped by (entityId, predicate); 2-fact groups are pairs the resolver left for adjudication, 3+-fact groups are multi-way disagreements escalated for human review. Answers "is this entity already conflicted on this predicate?" and can drive an in-product reviewer queue. asOf filters to disagreements that were live at that moment.

One entity: canonical name, type and what brain holds about it now; detail "full" adds externalRefs, the ids of its records in the originating vertical.

Chronological audit of all facts brain has learned about this entity, including retracted ones. Useful for "what did we know when" investigations.

Read a single fact as stored — predicate (aspect), object (statement), confidence, validity window, source attribution (vertical / recorder / conversationId), and lifecycle state. Retracted facts still resolve (retracted: true) — "why did I stop remembering this" is part of the trust story; only visibility fences turn into not-found. groundingStatus (grounded | ungrounded) appears on facts stamped by the claim-grounding plane (EVIDENCE_GROUNDING_STAMP); legacy rows carry no key. Same shape as GET /v1/facts/:id. Takes a factId as returned by searches and writes and returns the full trust record behind it.

Return the grounding provenance behind one fact: the verbatim conversation turns it was derived from (episodes, chronological, with char-span quotes when the deriver stamped them). When the server runs with the recursive-closure flag the response additionally carries derivedFacts + closure (the transitive derivedFrom support graph); a plain deployment returns episodes only — the response is a passthrough of GET /v1/facts/:id/provenance, no field filtering. Traversal depth is a server-side policy (flag + caps), not a caller knob. Lets a surprising fact be audited before it is trusted or retracted.

Return everything brain knows ABOUT a source (not its facts): the operator-declared identity (type, authLevel — authority in conflict resolution), the learned per-domain agreement rates (domain = predicate; null domain = the global blended rate; rates need >=8 samples before the resolver trusts them over the neutral 0.5), and the reputation-over-time history. sourceKey is `vertical:recorder` (recorder falls back to `_`) — the same key facts carry in trustSnapshot.sourceKey. Shows how much weight a fact whose trustSnapshot looks surprising deserves.

Resolves the named entities by canonical name, walks their 1-hop neighbourhood over knowledge_edge, and returns facts across seeds ∪ neighbours, optionally filtered by predicate hints. Works from entities named up front and returns what the graph holds around them ("who runs engineering at Acme" — the answer lives on a neighbour, not on Acme itself). Seeds score 1.0, neighbours 0.7; soft-fails to an empty result.

The two ways to authenticate, and what each is for. Reads nothing and stores nothing.

Insert an edge between two entities. `kind` is the edge type — `identity_of` merges the `from` entity into `to` (cross-vertical identity reconciliation), other typed edges (`paid_for`, `mentioned_in`, `worked_with`, …) are surfaced in related-entity reads and contribute to PPR / SubgraphRAG context. Most edges come from event ingestion; this inserts one directly. identity_of rejects a request that names the same end twice and any merge that would close a cycle; declaring an identity that already holds answers `alreadyIdentical: true` with no edge, so a replay is safe.

Paginated listing of topic communities, largest first. Each row carries a label, a rolled-up summary, and the member count. Gives the knowledge map at a glance.

Paginated listing of procedural memory for admin/review UIs. Sorted by priority ASC, createdAt DESC. Set includeRetired=true to see soft-deleted rows.

Cosine-matches procedural memory (curated 'how to' patterns the operator recorded) against a free-text context query. Returns top-K procedures sorted by similarity DESC then priority ASC. Surfaces the behaviour rules that apply to a context — e.g. "user asks about pricing" → "mention they're on platinum tier; they get 20% off". Procedural memory is the third tier alongside facts (semantic) and episodes (timeline).

Returns everything brain learned, unlearned, or replaced between two ISO 8601 cursors [from, to). createdFacts = new active facts; retractedFacts = facts marked retracted in-window with no successor; changedFacts = facts that were superseded by another (carries before+after); newEntities = entities created in-window; forgottenEntities = GDPR-erased tombstones; lapsedExpectations = temporary states (ill, away) whose expected end passed in-window with nothing confirming or ending them — worth asking about. Driving use case: "what changed since the last conversation?" Scope with entityIds and/or predicates to narrow the diff to a feature surface. Window is half-open; consecutive diffs over adjacent windows never double-count.

Find recorded design decisions / rationale / invariants / gotchas by NL topic ("why do we resolve facts through one gateway?"), across ALL code anchors — a semantic search over code-memory, rather than one known anchor. Returns matches with their code anchor + kind, ranked by relevance.

Capture the non-derivable engineering "why" about a code location: a design decision, its rationale, an invariant that must hold, or a gotcha. Anchored to a SCIP-style symbol or file path (NOT line numbers — survives edits). Stored as a bitemporal typed fact: re-recording a `decided` / `invariant` for the same anchor SUPERSEDES the prior one (decision evolution over time), while `because` / `gotcha` accumulate. Attach commit / location for provenance.

Records one fact about an entity directly; the memory settles it against the facts it holds (INSERTED / SUPERSEDED / COMPETING). Most facts arrive through ingestion.

Record a 'how to' pattern that is surfaced when a similar context appears later. trigger = the context phrase the rule should match against (e.g. "user asks about pricing"); action = the behaviour to apply (e.g. "mention they're on platinum tier; they get 20% off"). priority orders ties when multiple procedures match the same context (lower is higher priority; default 100). decayHalfLifeDays is a forward hook for v0.2 relevance decay; v1 ignores it at read time.

Mark a procedural memory row as retired (sets retiredAt). A retired procedure is no longer matched or listed by default. For a rule that no longer applies — distinct from a hard delete because the row stays for audit.

Mark a fact as no longer believed. Cascades to facts derived from this one. Does not delete; the row remains for audit.

Cosine-matches the query against topic-community summaries — the coarse retrieval scope. A community is a cluster of related entities (label propagation over the entity graph) with one rolled-up summary. Suited to broad questions ("what do we know about X domain"): it returns an overview of a domain rather than individual facts. Communities are (re)built off-hours by the dreams loop; an empty result means clustering has not run or the graph is too sparse.

Semantic search over the company knowledge graph: the facts the query is about, the links between their entities and what was left out. Apply asOf for historical "what was true on X" queries. Embedding model on this tenant: set per workspace.

Planner-LLM decomposes the query into ≤ maxHops anchored sub-queries; later hops are anchored to the running entity set. Built for questions that combine evidence across turns / sessions, or that require reasoning over multiple entities ("tenants who complained in April AND upgraded after"). Set synthesize=true to get a grounded answer with citations alongside the per-hop trace. Returns the hops, the answer when asked, and the evidence as focus lines. Embedding model on this tenant: set per workspace.

Returns a short one-line briefing about the entity — name, type, the most-confident active facts, external refs — suitable for dropping into an LLM context window. Caches in-process (per companyId / entityId / asOf / styleHint) so a hot entity touched across many turns doesn't reload the profile. styleHint='neutral' | 'sales' | 'support' are template-rendered (no LLM call). styleHint='client_llm' opts into MCP SAMPLING: brain asks the connected client (Claude Desktop / agent runtime) to write the one-liner with its own model — zero brain-side OpenAI cost, perfect for self-hosters who don't want brain holding an LLM key. Falls back to neutral template + sampledBy='local_template' when the client doesn't advertise sampling capability. A briefing in place of the full profile, timeline and competing facts.

Runs hybrid search then feeds the retrieved facts to a generator LLM that produces a citation-bearing answer (each claim ends with [factId]); a verifier LLM then judges whether every claim is supported. Three guardrail modes: strict (default) returns null on partial / unsupported / verifier outage (fail-closed); lenient returns the answer alongside the verifier verdict; off skips the verifier. Returns a direct natural-language answer rather than raw search results. Embedding model on this tenant: set per workspace.

Return the design decisions, rationale, invariants and gotchas recorded against a code anchor — the non-derivable engineering "why" a parser cannot recover from source. The anchor is a SCIP-style symbol string ("pkg/namespace/symbol") or a file path ("repo/path/file.ts"). Pass asOf to recall what was known at a past instant (bitemporal). Returns found:0 with empty memory when nothing is recorded.

The tenant this credential writes to (companyId, its display name, whether it is a personal workspace), what is already in its memory (entities, active facts, facts recorded in the last 7 days), how many Domain Packs are installed, for a user-bound connection who the user is in this memory (`user.name` — the name the memory has learned from them, null until it has), and `nextSteps` — the setup actions that would actually change something, derived from that state rather than a stored checklist (including asking the user how to address them, once, and recording it). Answers "which workspace am I writing to?" and explains why memory answers may look emptier than expected.