Spectron, under the hood

The full technical treatment of the Spectron memory engine: how conversations and documents compound into one governed company brain, how recall stays fast and cheap at scale, and how every answer can be defended when it reaches a stakeholder.

In this deep dive

01 · Memory model 02 · Document pipeline 03 · Hybrid retrieval 04 · Tiered queries 05 · Audit & time 06 · Autonomous understanding

Join waitlist Book a demo

Works with · +29 more

SOC 2 Type 2

GDPR

Cyber Essentials Plus

ISO 27001

01 |MEMORY MODEL

Six typed memory categories, not one bucket

Experiential memory is not a single store. It is six typed sub-stores - one raw record and five extracted categories built on top - each with its own schema, lifecycle, retrieval weight, and prompt-injection path.

Episodic

The raw conversational record. Sessions and turns as authored, in order, with anaphora intact. The source of truth every extracted category cites back to via byte-level spans.

Identity

Durable facts about the principal: name, role, employer, preferences, long-lived attributes. Long retention, low decay, surfaced into the prompt-ready profile.

Knowledge

Things the principal has learnt or shared, distinct from authoritative documents. Project facts, observations, references. Medium retention, decays without reinforcement.

Context

What is going on right now: active topics, recent intents, the working set for the current conversation. Short retention, replaced rapidly, anchors the next turn.

Instructions

Behavioural memory, not factual. 'Always answer in British English', 'never call me by my first name'. Applied at prompt-assembly time, not at retrieval.

Uncertainty

Explicit 'we don't know yet' rows. Emitted by the reconciler when confidence is below the floor, by cross-provenance contradictions, by ungrounded references. Gaps are visible, not papered over with hallucinations.

02 |DOCUMENT PIPELINE

Documents become structure - not a blob in a vector store

This is where the company brain gets built. Plain text, markdown, code, JSON, PDFs, images, audio and video flow through a proper knowledge-representation pipeline, landing in the same graph as everything your agents learn in conversation. Both the content index and the structural index that retrieval reads from are built at write time - so the cost ladder downstream stays tractable.

Multi-modal ingest

Configurable per Context via an IngestionProfile (TextOnly → MultimodalFull) that trades completeness against cost. Pay for what you actually need; lift it later without re-architecting.

Object store for originals

PDFs and media originals live in S3-compatible storage. The database holds the structured indexable state. Backups, GDPR deletion, and storage cost scale on object-store economics, not OLTP-row economics.

Content-addressed

Content addressing works exactly as in the ingest stage above - the Blake3 hash is the identity - and an extracted-text cache lets you rechunk and reembed without re-running expensive parsers.

Content-aware chunking

Passages are first-class chunk rows with their own embeddings, byte spans into the original artefact, and edges to the entities extracted from them. Time-coded segments for audio and video answer 'where in this recording' instead of 'this file is somehow relevant'.

Same reconciler as turns

Document-extracted facts flow through the same supersession-and-uncertainty function as turn-extracted facts. A document can contradict a turn and produce explicit uncertainty - identical guarantees, regardless of source.

Keyword graph + understanding

A non-LLM RAKE pass produces first-class keyword nodes with PMI-scored edges to chunks and entities - cheap structural recall for rare-term queries that vector search underweights. Document-level summaries, summary embeddings, and document-to-document links sit in the same graph.

03 |HYBRID RETRIEVAL

Eight signals fused into one auditable ranking

Embeddings-only retrieval has known failure modes - near-duplicates dominate top-k, rare-term queries miss the right chunk, structure between facts is invisible. Spectron's structural index is built at write time and read cheaply on every retrieval. Per-feature scores ride on every trace, so any result is auditable as a weighted combination of signals - not a black-box top-k.

The result is memory that stays coherent along five axes at once - semantic, lexical, relational, temporal, and spatial. Vector-only memory hits walls on four of the five; row-store memory hits walls on all five. The three clocks behind the temporal axis are covered in full under Audit and time below.

Vector recall

Dense embeddings on entities, attributes and chunks - plus modality-native vectors for images and audio when enabled.

Lexical (BM25)

BM25 over chunk text and entity surface forms catches the exact-phrase and rare-term queries embeddings underweight.

Graph traversal

One or two hops from a seed entity surface related facts even when surface forms diverge. The graph is the substrate, not a sidecar.

Keyword bridges

RAKE-derived keyword nodes connect chunks and entities that share rare or discriminative terms but sit far apart in embedding space. PMI-scored so common terms don't dominate.

Document links

Section embeddings, chunk-to-chunk and document-to-document edges surface related sections of related documents - not only the single closest chunk.

Personalised PageRank

Graph-walk scoring biased toward the query's seed nodes - a structural feature in the fused ranker, not a separate pipeline.

Geographic recall

Within-radius, inside-polygon and nearest-k spatial predicates compose with semantic and graph signals in the same ranker. 'Discussions about Acme within 50km of Berlin in Q3' is one query, not three pipelines.

Trace-derived features

Retrieval reads its own history. Rows useful for similar queries get boosted; rows associated with corrections get demoted. The ranker learns from the trace graph in place.

chunk

embeddingvector

text

extracted_from

attribute

text

confidence

valid_from

has_source

source

kind

trust

belongs_to

entity

type

name

embeddingvector

relation

kind (verb)·valid_from·valid_until

entity

type

name

embeddingvector

04 |TIERED QUERIES

Cheap questions are cheap. Expensive questions are explicit.

Most memory layers run a vector search on every request. Spectron does not. Reads route through a four-tier cost and latency ladder gated by query understanding at the front. Every tier emits a trace recording which path was taken, why, and what it returned - so the cost story is observable per Context, not buried in a flat per-request average.

Tier 1 · Direct lookup

Typed questions - 'what is my role at Acme?', 'when did I join?' - fetched directly from the entity and attribute graph by key. No embeddings, no LLM, no ranking. Sub-millisecond.

Tier 2 · Response reuse

Semantic match against prior responses with entity-aware invalidation. The cache key is the query plus the set of facts the prior answer cited - when any cited fact is superseded, every dependent response is invalidated. Tens of milliseconds. No LLM call.

Tier 3 · Hybrid retrieval

Vector, BM25, graph traversal, keyword bridges, section embeddings, Personalised PageRank, geographic recall and trace-derived features fused into a single ranking, then synthesised by the LLM. Hundreds of milliseconds. Auditable per-feature scores ride on every retrieval trace.

Tier 4 · Full-context fallback

When hybrid retrieval is thin or low-confidence, fall back to a broader sweep: more candidates, deeper graph traversal, optionally a HyDE-style query rewrite, longer context window. Higher token cost and latency - explicit as the tier-4 escalation, not buried in a flat per-request average.

Natural language query

Generate embeddingsProvider-pluggable

Vector searchchunks

HNSW · COSINE

Entity searchentities

HNSW · COSINE

Fact chain traversalentity → relation → entity

3-depth recursive DFS

Memories + facts

05 |AUDIT & TIME

Every fact knows where it came from - and when it was true

The risk teams weigh most with agentic AI is a confident wrong answer reaching a stakeholder. Spectron's answer is structural: provenance is a stored field on every fact-bearing row, three independent clocks answer three different questions, and traces are first-class nodes that feed back into ranking. The audit story is queryable substrate state - not a logging pipeline bolted on.

Provenance as data

Every entity, attribute, relation, instruction, and uncertainty row carries a source object: kind (turn, document, reflection, elaboration, consolidation), ref, trust, lexical span, location, and derived_from. Reconciliation compares provenance; calibration weights it; supersession audits it - down to the bytes.

Three independent clocks

System time (the MVCC byte-level audit clock - 'what did the substrate look like at instant T?'), known time (when Spectron first believed a fact), and valid time (when the assertion held in the world). Different questions, all queryable independently via VERSION and as_of.

Traces are memory

retrieval_trace, decision_trace, and response_trace are first-class graph nodes, not external observability data. The unified ranker reads its own history; supersession lineage downgrades source trust; any answer walks back to the bytes that produced it.

Supersede, don't delete

Memories are not overwritten - they are superseded with valid_until set, or aged out with a reason recorded. The history of how a fact changed is queryable. forget is an explicit first-class verb, distinct from natural aging, with a --purge option for hard removal.

06 |AUTONOMOUS UNDERSTANDING

Memory that improves between conversations

Three named mechanisms generate new memory at different times, all flowing through the same reconciler with explicit provenance. Spectron does not just store what agents tell it - the substrate deepens its own understanding between interactions, so the tribal knowledge tenured staff feed it compounds into shared memory every teammate's agent can recall.

Reflection

On-demand synthesis. POST /reflect runs an LLM pass over retrieved context and optionally persists the synthesised answer as new facts - with their own provenance kind and a lower default trust, so calibration stays honest.

Elaboration

Background sweep. A job walks the substrate looking for entities and attributes that share context but no explicit relation. An LLM proposes the link; the reconciler accepts, supersedes, or surfaces it as uncertainty.

Consolidation

Belief crystallisation. An async job pools recent facts and decides to create, update (delta recorded), or mark superseded. Each observation tracks its derived inputs and proof count, so the evolution of belief is replayable.

GET STARTED

Be first to build with Spectron

Join the waitlist for early access - every self-serve plan starts with a one-week free trial at launch.

SOC 2 Type 2

GDPR

Cyber Essentials Plus

ISO 27001

Join waitlist Back to the overview