01 / PROVENANCE
Every important state has a lineage.
Witness records bind observations, claims, transformations, hashes, and review context so future reasoning can distinguish evidence from assumption.
Eidolon Quantum Systems, Inc.
Memory that can
show its work.
We are developing Anamnesis Rising: North Star — governed memory and semantic infrastructure for AI systems that need continuity, contradiction awareness, replay, provenance, and explicit authority boundaries.
Active technical preview · local evidence first · no hidden promotion
synthetic semantic field / illustrative
The design problem
Agents need memory. Memory needs governance.
A long-running AI system is not helped by merely remembering more. It needs to know where a memory came from, what contradicted it, whether it was only a candidate interpretation, who authorized any durable transition, and whether the path can be replayed.
02 / CONTRADICTION
Disagreement is retained, not paved over.
North Star preserves counterevidence and unresolved tension instead of forcing the system toward a single convenient narrative.
03 / AUTHORITY
Recall is not permission.
A memory hit, ranking score, semantic match, or model conclusion does not by itself authorize action or durable mutation.
Perception-first direction
Observe → interpret → compare → witness → review.
“Perception-first” is the design direction: start from observed signals, context, uncertainty, provenance, and contradiction before treating a model output as a command. It is an operating architecture goal, not a claim that North Star is already a general-purpose production OS.
SignalConversation, event, sensor or structured input.
Candidate meaningInterpretation remains bounded and attributable.
Spatial contextQuery semantic neighborhoods and relationships.
WitnessRecord what was seen and how it changed.
ChallengeCompare countermodels and contradictions.
Governed transitionOnly explicit authority may make durable change.
Core architecture
Four systems working together.
North Star treats memory as an evidence-bearing systems problem: semantic location, differentiated interpretation, replayable state change, and controlled authority are designed as separate concerns.
SPATIAL MEMORY
Memory as navigable semantic terrain
Serialized records are interpreted spatially at runtime, allowing neighborhood queries, relationship-based recall, contradiction separation, and replay without pretending the disk literally stores a “memory universe.”
DUAL-STATE
Independent interpretations
One state expands possible meaning; another challenges it from evidence, provenance, consequence, and continuity. Agreement is never forced.
ANLS
Semantic packets, not just text blobs
Templates, coordinates, slots, deltas, lineage, hashes, and reconstruction receipts provide a compact structured representation while keeping boundary text reconstructable.
WITNESS + GOVERNANCE
Explain behavior change without granting memory authority
Replay, Memory Diff, review queues, and append-only receipts make the system accountable while preserving a hard separation between evidence and permission.
ANLS benchmark · Phase 82 synthetic fixture
46.98% fewer bytes than literal JSON in the tested aggregate corpus.
That is useful — but it is not a “compression beats compression” story. Deterministic zlib level 9 was substantially smaller. ANLS is interesting because the representation carries semantic coordinates, dictionaries, lineage, deltas and exact reconstruction semantics while still reducing storage relative to literal JSON in this fixture.
Open full benchmark80 project-generated synthetic records. Exact byte accounting. Not a latency, retrieval quality, safety, truth, or production-performance benchmark.
Engineering principle
Give agents memory without giving memory the keys to rewrite reality.
North Star is being built so a future reviewer can reconstruct what the system remembered, where it came from, what contradicted it, how behavior changed, and who authorized each durable transition.