Architecture

Spectron glossary

Alphabetical definitions of terms used across Spectron architecture and memory docs.

Short definitions for terms that appear throughout the Spectron docs. For deeper treatment, follow the links where provided.

Spectron’s vocabulary is precise, but the intent is everyday: memory that associates related ideas (like hearing “cat” and thinking of pets, lions, or a nickname) and keeps time and scope straight (“I have one” versus “I used to” versus “they weigh about four kilos in general”). The entries below name the machinery; What is Spectron? and Memory categories walk through the human-shaped examples.

Anaphora — Pronouns and phrases in conversation that refer back to something said earlier (“he”, “that project”, “the policy we discussed”). Episodic memory keeps the transcript intact; extraction uses context to resolve what they mean.

Attribute — A named property on an entity (for example role = Head of Platform). Attributes can be superseded over time with valid_from / valid_until.

Authoritative (pillar) — Curated, vetted knowledge: handbooks, policies, product catalogues. Usually ingested as documents with higher default trust.

Authority — The policy that curated sources win over casual chat when sources conflict, implemented by the reconciler (often via uncertainty rather than silent overwrite).

Calibration (pillar) — Gating assertions by confidence and trust; below a floor, Spectron refuses to overwrite and records uncertainty instead.

Collective (pillar) — Shared memory when multiple principals corroborate the same fact; promotion to wider scope is policy-gated (automated promotion not yet enabled in early preview).

Confidence — The reconciler’s posterior certainty on an assertion (attribute.confidence), stored separately from source trust.

Context (read parameters) — Optional labels, lens, and scope_view on /query and /context to filter or fold results within what your grant already allows. See Contexts and scope.

Consolidation (pillar) — Background process that crystallises recurring observations into longer-lived beliefs.

Context (category) — Short-lived “what matters right now” for the active conversation.

Context (tenant) — A Context is Spectron’s top-level isolation unit: its own database, keys, and configuration. Not the same as the memory category “context”.

Decision trace — A graph record of what the reconciler considered, created, or superseded during a write.

Delegation — Acting on behalf of another principal via X-Spectron-On-Behalf-Of, intersected with both parties’ grants (depth 1).

Elaboration (pillar) — Background linking of related facts that were stored separately but belong together.

Entity — A typed node in the memory graph (for example Person/alice, Product/airpods_pro).

Episodic (category) — The raw ordered transcript of sessions and turns before extraction.

Experiential (pillar) — Knowledge from conversation and derived passes (turn, reflect, elaboration, …), usually lower default trust than authoritative material.

Extraction — The pipeline that turns text into entities, attributes, relations, instructions, and uncertainties.

Forget — Explicit soft-removal of matching memory (valid_until), distinct from aging.

Identity (category) — Stable facts about who the principal is (name, role, employer).

Instructions (category) — Behavioural preferences (“use British English”), applied at prompt assembly, not generic retrieval.

Graph edge kind — A named structural signal used in document hybrid_graph retrieval (for example knowledge_has_keyword, document_link). Request and response enums reject unknown values with 400 Bad Request.

Known time — When Spectron first recorded a belief; queried with as_of. See Tri-temporal model.

Knowledge (category) — Facts learnt in conversation, distinct from uploaded manuals.

Memory category — On extracted entities, attributes, and relations: identity, knowledge, or context (API-enforced enum).

Principal — An actor (human, agent, service) that can be granted read/write permissions on scope paths.

Provenance — Structured metadata on every fact: source kind, reference, span, trust, derivation. See Provenance and traceability.

Reconciliation — Comparing new extractions to existing memory: supersede, merge, or emit uncertainty.

Reflection (pillar) — On-demand synthesis from a question; results stored with lower default trust.

Response trace — Record of a composed answer (/chat, /reflect), used for tier-2 reuse when still valid.

Retrieval trace — Record of a /query (or related read): candidates, signals, tier, and results.

Scope — Tags on records (org=acme, user=alice) that partition data within a Context. Subset matching controls what a query can see. Distinct from permissions.

Session — A conversation container: ordered turns, scope, and metadata.

Source.kind — Whether a fact came from a document, turn, reflect, elaboration, consolidation, and so on.

Source.span — Character positions in the originating message or passage for quotable citations.

Source.trust — Default weight reflecting how authoritative the source stream is (document versus casual turn, and so on).

Supersession — Replacing a prior assertion by setting valid_until on the old record while keeping history.

System time — Database MVCC history: what was stored at an instant. See Tri-temporal model.

Trace layer — Memory about how memory was used: retrieval, decision, and response traces.

Trust — See source.trust and calibration; central to why curated and conversational facts combine safely.

Turn — One message in a session (user, assistant, system, tool).

Tri-temporal — System, known, and valid time kept distinct. See Tri-temporal model.

Uncertainty (category) — Explicit “we do not know” or conflict records instead of guessed fill-ins.

Unified graph — One SurrealDB store for documents, turns, entities, relations, and traces. See Unified substrate and authority.

Valid time — When a fact held in the real world (valid_from / valid_until), independent of when Spectron learnt it.

Was this page helpful?