Current and as-of truth
Ask what is believed now, what was believed at a past recording time, or what was valid during a historical interval.
SolarflareDB Memory
Persistent agent memory that knows when.
A memory should not be a detached text chunk. SolarflareDB treats each result as a temporally scoped, source-backed part of a living context graph.
Shared memory with identity
Associate writes and retrieval decisions with tenant, agent, session, request, source, and authorization context.
History that remains explainable
Supersession and retraction update current interpretation without erasing the claims and episodes that produced it.
Memory lifecycle
Ingest, resolve, retrieve, and explain.
Agent memory quality depends on more than storage. The product must make extraction, entity resolution, temporal interpretation, retrieval, context assembly, and feedback observable.
Observeepisode + source + event time
→
Resolveentities + facts + valid time
→
Retrievehybrid signals + policy + provenance
Built for difficult cases
Where ordinary vector memory begins to blur.
SolarflareDB’s wedge is the moment a team needs context to change safely, be shared across agents, or answer a question whose evidence spans time and relationships.
Entity resolution without irreversible collapse
Normalize exact identifiers, generate fuzzy and embedding candidates, maintain aliases, and record reversible merge decisions with confidence.
Contradiction and supersession
Model “no longer true,” “was incorrect,” and “true only during this interval” as explicit context relationships.
Immediate post-write retrieval
Union derived index candidates with a bounded authoritative recent-write overlay until lexical and vector watermarks catch up.
Explainable authorization
Return only context admitted by tenant, graph, subject, agent, and source policy—and preserve the decision metadata.
Context assembly
Optimize useful context, not retrieved tokens.
A query may find hundreds of candidates. The assembly layer deduplicates, applies temporal and policy filters, diversifies evidence, reranks, and selects a context set that fits the model budget.
Exact and identifier candidateshigh precision
Lexical and semantic candidatesrecall
Graph-neighborhood candidatesmulti-hop
Temporal and policy interpretationtruth
Budget-aware context setanswer
Give one agent a memory. Then let another trust it.
The simulated console shows the intended developer experience without pretending a production backend is already connected.