While auditing v0.5.0 as a reference implementation, I noticed the README and a few docs describe shapes that don't match the shipping code. Flagging in case you want to bring them back in sync — the drift was significant enough that a README-only first pass on my side had to be substantially retracted after reading the code.
1. Compression label vocabulary. README's "Compression Labels" section lists 5 labels (CRITICAL / DEBUG / TOOL_OUTPUT / ERROR / INFO). hive/rule_fast/__init__.py defines 6 (CORE / DISTILL / COMPACT / DROP / STALE / ESCALATE), and the module docstring explicitly references the 6-label vocab as honey-comb's authoritative taxonomy. README appears to predate the code.
2. Vector / semantic search. README's API section implies Hive is keyword-only (recall(key) -> List[MemoryNode]). hive/semantic_search.py ships a SemanticSearch class doing cosine over sentence-transformers embeddings — a non-trivial capability the README doesn't mention.
3. Multi-tenancy. README presents HiveStack as in-process single-tenant. RustBrain(tenant_id=..., tenant_isolation=True) prefix-isolates per tenant, and revoke_tenant() mass-erases (per the GDPR Article 17 comment). Harness docs reference these; README doesn't.
4. Durability. README is silent on persistence. Code has snapshot_to_file() (gzip + SHA-256) and restore_from_file(). Worth a README mention since "will my data survive a restart" is a common evaluator question.
5. Rust crate (hive-cpp) status. README's perf table cites "100×" speedup vs Python. The current hive-cpp/src/memory.rs is a HashMap in OnceLock with max_capacity: 10_000 — paralleling the Python reference rather than the SIMD/roaring-bitmap target described in the rust_brain/__init__.py docstring. A "current vs roadmap" note in the README would set expectations correctly.
Filed to help; not a blocker. Thanks for the open-source work.
While auditing v0.5.0 as a reference implementation, I noticed the README and a few docs describe shapes that don't match the shipping code. Flagging in case you want to bring them back in sync — the drift was significant enough that a README-only first pass on my side had to be substantially retracted after reading the code.
1. Compression label vocabulary. README's "Compression Labels" section lists 5 labels (
CRITICAL/DEBUG/TOOL_OUTPUT/ERROR/INFO).hive/rule_fast/__init__.pydefines 6 (CORE/DISTILL/COMPACT/DROP/STALE/ESCALATE), and the module docstring explicitly references the 6-label vocab as honey-comb's authoritative taxonomy. README appears to predate the code.2. Vector / semantic search. README's API section implies Hive is keyword-only (
recall(key) -> List[MemoryNode]).hive/semantic_search.pyships aSemanticSearchclass doing cosine oversentence-transformersembeddings — a non-trivial capability the README doesn't mention.3. Multi-tenancy. README presents
HiveStackas in-process single-tenant.RustBrain(tenant_id=..., tenant_isolation=True)prefix-isolates per tenant, andrevoke_tenant()mass-erases (per the GDPR Article 17 comment). Harness docs reference these; README doesn't.4. Durability. README is silent on persistence. Code has
snapshot_to_file()(gzip + SHA-256) andrestore_from_file(). Worth a README mention since "will my data survive a restart" is a common evaluator question.5. Rust crate (hive-cpp) status. README's perf table cites "100×" speedup vs Python. The current
hive-cpp/src/memory.rsis aHashMapinOnceLockwithmax_capacity: 10_000— paralleling the Python reference rather than the SIMD/roaring-bitmap target described in therust_brain/__init__.pydocstring. A "current vs roadmap" note in the README would set expectations correctly.Filed to help; not a blocker. Thanks for the open-source work.