Knowledge that knows who's asking.

Pipeline, where agents build it

• Stages: triage, render, extract, validate, commit, truth maintenance.
• Triage and budgets (agent): a model picks each source's depth, skip, metadata, retrieval or graph, daily budgets can lower the depth, cost reserved before the work starts.
• Extraction (agent): an agent on any model through Meerkat, reads Meerkat Recall over MCP as it extracts, checks evidence against the text, strict structured output, stages candidates for commit.
• Staged commit: a durable work queue with leases, validate, then commit, retries with backoff, then dead letters, idempotency keys on ingest and claims.
• Truth maintenance (agent): runs after each commit, agreeing values merge their evidence, an LLM adjudicates disagreements, supersedes by stated validity dates, proposes semantics for new predicates.

Knowledge model, what agents query

• For agents: MCP (/mcp), MCP agent (/mcp/agent), HTTP API (/v1), Event stream (SSE), Rust crate (library); tools grouped into MCP profiles, a read-only agent profile, the extraction agent reads through it too, get_runtime_status on every profile, ingest tools hidden until extraction is ready.
• Sources and revisions: nine source types, email to video, every render an immutable revision, images and video rendered to text.
• Documents and layered embeddings: a tree of nodes with byte-offset spans, chunk, paragraph and summary vectors, vector search, lexical fallback.
• Assertions and evidence spans: what it was told, kept as evidence, subject-predicate-object claims, byte offsets into a document revision, a quote hash on every span.
• Entities and identity (agent): names, aliases, handles, identity keys, type-guarded identity resolution, an LLM settles ambiguous matches, an identity sweep proposes merges, merges keep lineage and redirects.
• Relationships and attributes: entity to entity, or to a typed value, predicates from a governed registry, validity windows, confidence, evidence.
• Truth slots and conflicts: what it believes, one answer per question, ranked values with a rationale, conflict groups keep competing claims, beliefs never mix security envelopes.
• Timeline: events backed by claims and sources, precision from exact to year, or unknown, later evidence refines the same event.
• Security: policy, level, labels, handling; designed in from the start, an envelope on every record, deny by default, eight access rules, JWT principals, clearances per policy, derived records inherit the envelope, vectors partitioned by policy, level, model, fields redacted without the scope.

Storage, the same code on either backend

• SurrealDB: persistent, the default backend, document, graph and vector features, HNSW indexes per vector partition.
• ElephantDB: embedded kernel, in the same process, exact cosine vector search, snapshot export and import.