SPECTRON
One substrate.
One transaction.
One memory layer.
Documents, conversations, entities, attributes, relations, embeddings and traces all live in one ACID-transactional database. Spectron is the stateless application tier on top - no cross-store stitching, no consistency gaps between vector and graph, no sidecars to operate.
01 |THE PROBLEM
Memory wrong is worse than no memory
Most agent memory is a vector index stitched to a graph store and a row store, with the seams smoothed over by application code. It works in demos and breaks in week three. Four failure modes are predictable - and none of them are retrieval problems.
02 |EIGHT PILLARS
The model that holds it together
Eight primitives the substrate, the write path, and the read path all have to support. Not features added on top - the only eight things memory has to do to hold up under scale, change and multi-instance deployment. Everything else in Spectron is a way to make one of them work in production.
03 |HOW IT WORKS
Ingest, extract, connect, query
Four stages turn conversations and documents into a typed entity and relation graph - with provenance on every row, supersession instead of overwrite, and a trace recorded for every read and write.
FIVE COHERENCE DIMENSIONS
Memory that holds up along five axes
Vector-only memory hits walls on four of these five; row-store memory hits walls on all five. Spectron's substrate stores enough metadata to answer questions on each axis - simultaneously, in the same fused ranker.
04 |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.
05 |DOCUMENT PIPELINE
Documents become structure - not a blob in a vector store
Plain text, markdown, code, JSON, CSV, PDFs, images, audio and video flow through a proper knowledge-representation pipeline. Both the content index and the structural index that retrieval reads from are built at write time - so the cost ladder downstream stays tractable.
06 |AUDIT & TIME
Every fact knows where it came from - and when it was true
Provenance is a stored field on every fact-bearing row. Three independent clocks answer three different questions. Traces are first-class nodes that feed back into ranking. The audit story is queryable substrate state - not a logging pipeline bolted on.
07 |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.
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.
08 |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.
09 |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.
10 |IN PRODUCTION
Built for the second deployment, not the first demo
Calibration, scope isolation, observability, and security are stored substrate state - not configuration claimed in a README. Auditors verify the same way operators debug: by reading the graph.
MODEL CONFIGURATION
Five model hooks, mixed and matched per Context
The right model for extraction is rarely the right model for synthesis, and neither is the right model for embedding. Each hook is an independent knob - configurable per Context, overridable per call, and recorded on every trace - so cost and quality slice per model out of the box. Anything OpenAI-compatible, Anthropic, Google, or local inference. Air-gapped deployments stay air-gapped.
INTEGRATIONS
MCP, generated SDKs, and harness adapters
One OpenAPI specification is the source of truth. The Spectron binary speaks Model Context Protocol natively, ships generated clients in four languages, and offers thin adapters for the harnesses that don't natively speak MCP - so clients cannot drift from the server.
11 |BUILT ON SURREALDB
Multi-model, in one ACID transaction
SurrealDB unifies graph, vector, document, relational, and geospatial queries in one engine. The eight pillars, six memory categories, five coherence dimensions, and the trace graph are all expressible as rows and edges in a single transaction - which is what makes them deliverable as a coherent product instead of a stitching exercise.
Multi-model ACID
Entities, attributes, relations, embeddings, chunks, and trace edges commit atomically. No cross-store stitching, no eventual consistency between vector and graph.
Free time-travel
SurrealDB's MVCC layer means SELECT ... VERSION returns the exact substrate state at any past instant - the byte-level audit clock is a database feature, not application code.
Record-level permissions
Each Context is its own namespace and database; per-principal grants, RBAC, and per-record rules apply to memory the same way they apply to application data.
Scale-to-zero substrate
Compute-storage separation means an idle Spectron deployment costs nothing - no minimum cluster, no always-on vector index to pay for between conversations.
Memory branching
Compute-storage separation will let you branch an entire knowledge graph in seconds - a copy-on-write clone of the whole substrate for testing or evaluation, not an export stitched back together across separate stores.
12 |ARCHITECTURE
Middleware on fragments, or memory in the database
Most memory layers stitch two or three stores together - Postgres with pgvector and Neo4j, or DynamoDB with OpenSearch and Pinecone - and inherit the seams: no cross-store transactions, divergent consistency models, separate scaling stories. Memory middleware abstracts over the fragmentation but cannot eliminate it. Spectron removes it.
13 |WORKED EXAMPLE
A user changes their mind. Spectron tells the truth.
Three turns of a real conversation. The old fact is preserved, the new fact supersedes it, and the answer comes back with the trace that proves it.
TURN 1 · 15 MAR 2026
User: "I live in Berlin."
TURN 7 · 12 APR 2026
User: "Actually I moved to Paris last month."
QUERY · LATER
Agent: "Where does Emma live?"
FOR YOUR TEAM
Two paths into Spectron
One product, two on-ramps. Developers and AI engineers join the waitlist for early preview access; engineering and platform leaders evaluate fit for production with the team.
For developers and AI engineers
One static Rust binary with the MCP server built in - connecting Cursor, Claude Desktop or Claude Code is a single configuration entry. Native SDKs in Python, TypeScript, Kotlin, and Swift, generated from the OpenAPI spec. No Python in the pipeline, no sidecars. Join the waitlist for early preview access.
For engineering leaders and platform teams
ACID writes across graph, vector, document and structured records. Context-level tenant isolation enforced at the engine. Provenance, tri-temporal history, and trace-graph audit baked into every fact. Self-host on SurrealDB, run on SurrealDB Cloud, or deploy in an air-gapped environment with local model inference.
14 |MEASURED · OPEN
Numbers we will publish. Source you can read.
Trust is what gets agents into production. Two ways we earn it: the benchmarks we measure against, and the open-source database every Spectron deployment runs on.
Measured against published benchmarks
Spectron is evaluated on LoCoMo and LongMemEval, the two published conversational-memory benchmarks, alongside StateBench, an in-tree state-tracking suite, plus a document-retrieval regression harness with a factoid-versus-graph reporting split. Each run produces a comparable report against any commit, so ranking and reconciliation changes ship with measured deltas, not anecdotes.
Open at the foundation
SurrealDB, the database engine underneath Spectron, is open source and free to self-host (32.3k GitHub stars and counting). The Spectron memory layer on top is closed source today and ships as a single Rust binary. Our roadmap intent is to upstream foundational parts of the memory model into SurrealDB, so the most fundamental primitives stay open.
FREQUENTLY ASKED QUESTIONS
