ANLS semantic storage

Represent meaning as structured state — then prove you can reconstruct the boundary view.

ANLS is North Star’s native semantic representation research line. Instead of treating every memory as an opaque text blob, an ANLS packet can carry a versioned template, semantic coordinate, slots, deltas, lineage, privacy classification and stable hashes, with a reconstruction receipt binding the packet to its rendered form.

Warm Indexversioned templates + coordinate families
+
ANLS Packettemplate · slots · deltas · lineage · hash
Boundary rendererdeterministic reconstruction
Receiptpacket/index/body hashes bound together
Why it matters

Compression is only one part of the value proposition.

The key research question is whether a semantic representation can remain compact while preserving structure needed for retrieval, lineage, contradiction, deterministic rendering and governance. ANLS is not positioned as a replacement for mature byte-compression codecs.

STRUCTURE

Meaning is addressable

Templates, slots and semantic classes can make recurring structure explicit instead of re-encoding every record as full literal prose.

LINEAGE

Context travels with the packet

Representation metadata includes provenance and version bindings needed to explain where a semantic state came from.

RECONSTRUCTION

Boundary text stays verifiable

Receipts bind packet, index version and rendered body so a compact internal representation can still be checked against its human-readable view.

Deterministic synthetic benchmark

80 records. Five representations. Exact byte accounting.

The Phase 82 fixture was project-generated and synthetic: 64 repeat-heavy records plus 16 heterogeneous records. Every representation counted the bytes required for exact reconstruction.

ANLS aggregate
14,833 B
46.9777% savings vs literal
zlib-9 aggregate
2,956 B
89.4334% savings vs literal
Literal JSON
27,975 B
JSONL
27,925 B
Structural
27,272 B
ANLS
14,833 B
zlib-9
2,956 B
CohortLiteralzlib-9StructuralANLSANLS savings
Repeat-heavy (64)23,067 B1,897 B22,108 B10,786 B53.2406%
Heterogeneous (16)4,958 B1,310 B5,228 B4,105 B17.2045%
Aggregate (80)27,975 B2,956 B27,272 B14,833 B46.9777%
Important negative result: zlib was the smallest representation in every fixture cohort. The benchmark does not prove universal compression superiority, retrieval quality, semantic utility, latency, safety, truth, or production performance. On heterogeneous data, the structural-template representation was actually larger than literal JSON.
Research advantage

Why pursue ANLS if zlib is smaller?

Because byte size and semantic usability answer different questions. A conventional compressor can minimize bytes extremely well. ANLS explores whether the stored representation itself can expose stable semantic slots, coordinates, lineage and deltas that downstream memory/replay systems can reason about directly.

Queryable structure

The representation can expose semantic fields without first pretending decompressed prose is the only canonical computational form.

Versioned semantics

Warm-index and template versions make interpretation dependencies explicit and auditable.

Governed translation

Translation/reconstruction receipts preserve authority-null semantics across representations instead of smuggling permission through format conversion.