These are the release notes for every version, newest first, generated from the commit history.

Changelog¶

All notable changes to this project are generated from the commit history. Format: Keep a Changelog + Conventional Commits.

[Unreleased]¶

Added¶

• One series per patience value in the ANN comparison tables
• Stemmed BM25 and min-max score blend in the document API
• Patience stop rule, opt-in 1-bit screen and error-returning builds for the vertex graph index
• Competitor ANN harness for SymphonyQG, RaBitQ, NGT and Lucene HNSW
• Check the graph cache header and verify the merged qgraph defaults
• Add integer vector stores and per-vertex codes to qgraph
• Batch Qwen3 decoding and padded encoder batches on Metal and CUDA
• Add per-vertex 8-bit code graph index with residual re-rank
• Add held-out tuning mode to public-bench ann-qgraph
• Add on-device Qwen3 model stages behind local-models
• Add quantized graph ANN index and its public benchmark mode
• Add hms-server HTTP service for the Python SDK
• Add experimental query-private similarity scoring behind private-search
• Subset-union bundle experiment and capacity sweep

Changed¶

• Drop rejected intersection kernel candidates

Documentation¶

• Ideas tree for October 2026 with its 10-minute proxies
• Preliminary nytimes VGraph comparison on the M4 and the next-session specs
• Record the green CI and the clippy toolchain for the gate
• Record that main is pushed and in sync
• Record that all work is on main and what was superseded
• Record the merged qgraph state and what is unverified
• Record the integration state
• Standardize README header (#71)
• Update continuation state
• Record the index and model-stage outcomes
• Record the holographic prototype outcome
• Record continuation state for the next session
• Plan the holographic fact memory and its tests
• Plan local model stages inside HMS
• Define the single-run protocol for LongMemEval_M
• Set the speed and accuracy targets and record the compute limits
• Plan the LongMemEval_M run with larger models
• Record the encoder test and the agent-memory benchmark plan
• Attribute semantic search to the document API
• Update changelog [skip ci]
• Record the outcomes of the algorithm work
• Give the SDK README the project header, badges and a quick start
• Use absolute logo URLs so the README renders on package registries
• Update changelog [skip ci]

Fixed¶

• Satisfy clippy 1.99 lints in the qgraph kernels and graph cache
• Keep greedy decoding at batch 1 unless a batch is set
• Reserve Modal wave cost before launch so preemption cannot bypass the cap
• Run HMS model stages on L4 on Modal
• Decode benchmark data with as_chunks for the current clippy
• Resolve leftover merge markers in .bestpractices.json
• Serve SDK queries by exact cosine through the document API
• Return query results best-first on exact-scan, multi-shard and federated paths

Performance¶

• Quantized construction search for the graph build; cache header records it
• Quantized construction search and owned-input build for the proximity graph
• Byte visited marks for the unfused block path; close the neighbour-block mechanism
• Keep the default neighbour loop free of the delta decoder and the cap counter
• Neighbour cap for the per-row search, lint fixes, x86 paired rows
• Unbiasing scale for block codes, sorted candidate pool, delta-coded adjacency
• Neighbour-block fast-scan codes with evidence fusion and a certified exit for the vertex graph
• NEON per-row 1-bit screen kernel with its microbenchmark kill test (screen stays off)
• Scan both planes of 2-bit qgraph edge codes in one NEON pass
• Add FastScan edge kernel, batched pool merge and 2-bit qgraph edge codes
• Add 4-bit traversal codes, aligned rows, BFS reorder and 3-byte ids to the vertex index
• Screen exact distances in vertex graph construction with 8-bit bounds
• Measure Metal batching and set per-stage defaults
• Prefetch rows ahead in qgraph fill, prune inputs and edge encoding
• Use a bitset visited set and prefetch adjacency in qgraph construction
• Speed up qgraph construction without changing the built index
• Speed up from_dense with branchless term compaction and add encoder-eval
• Gallop before AVX2 on skewed sizes; add intersection kernel bench
• Route sparse queries to the exact inverted index

Bench¶

• Per-session quota on the tuned LongMemEval turn ranking fails on cached dev scores
• Record timed held-out glove run of i8 vertex-code qgraph
• Correct reference eval counts in memory-first results
• Record held-out recall screening of qgraph memory layouts
• Add grid-matched vertex search pairs and correct the alignment claim
• Record paired held-out timing of the vertex search encodings
• Narrow the FAISS comparison and vendor the clean edge baselines
• Record paired held-out timing of the per-vertex code index
• Add paired-timing sync to ann-qgraph
• Record held-out recall and memory of the per-vertex code index
• Record the s-dev-hms cost overrun and stop
• Record held-out qgraph tuning on nytimes
• Measure HMS fact extraction against the Python stage on S dev
• Add seeds and an equal-bytes index control to the holographic prototype
• Record quantized graph results against HNSW
• Prototype the holographic fact memory and measure it on LongMemEval_S dev
• Make the Modal cost cap cumulative across relaunches
• Key LongMemEval fact extraction by content and re-tune on dev
• Record the LongMemEval dev results and the Modal dry run
• Tune the LongMemEval pipeline on dev and record the ablation
• Add the LongMemEval multi-key pipeline, Modal job and dev sweep
• Freeze the LongMemEval 100 dev / 400 held-out split
• Add LongMemEval_S retrieval baseline (dense, lexical, hybrid)
• Add public BEIR and ann-benchmarks comparisons
• Report encoder candidate recall after exact rerank

Research¶

• Durability matrix harness and first matrix against SQLite and Lucene
• Record native benchmark gate checkpoint
• Checkpoint native harness validation draft
• Establish checked LongMemEval dev answer results
• Checkpoint final local reader batch and faster validation
• Preserve dev reader progress and speed pilot audits
• Checkpoint local reader evidence and judge progress
• Checkpoint judge instability and local evidence
• Checkpoint dev answer scores and repeat study
• Freeze dev reader protocol and checkpoint quota blocker

[0.6.1] - 2026-10-05¶

Documentation¶

• Update changelog [skip ci]

[py-v0.6.1] - 2026-10-05¶

Added¶

• Introduce IdentityRegistry enforcing W3C DIDs, Verifiable Credentials, and POSME Receipts for queries
• Add AsyncHolographicClient, batch uploading, and metadata filtering
• Add Pydantic AI and SmolAgents tool wrappers
• Add Phidata integration and fix broken Rust examples
• Add Semantic Kernel and DSPy integrations to Holographic SDK
• Add LlamaIndex SDK adapter and CI pipeline
• Pivot to Zero-Trust Edge Architecture (FHE-Lite, AGPL-3.0, SIMD)
• Python bindings + PyPI packaging for holographic-vsa (maturin/pyo3)
• PhaseResonator reusable-index API + clean re-export + doctest (§20 capability)
• PhaseHVec core type + real Frady/Kent resonator on quantized-phase substrate
• Counting membership store with Poisson z-score readout (4x Bloom capacity)
• Real SHA-256 commitment for verifiable phasor mutation (lane 3)
• Multi-hop reasoning over the phasor memory (retrieve_path)
• PhaseGraph -- phasor relational memory with relation algebra + retrieval
• Presence-gate the observer effect (no confabulation-on-query)
• Persist ConnectionGraph via its event log; survives restart
• ConnectionGraph -- plastic event-sourced relation store, wired into engine
• Trust-anchored provenance verification and sharded ANN persistence
• Runnable cross-verification of the cogmem C2PA sample
• Add provenance system with COSE, VCs, C2PA, JUMBF, Sigstore, KERI, CAWG
• Add research-grade benchmarks, scaling analysis, and visualizations
• Add sparse Clifford algebra multivector type (CliffordVec)
• Define HolographicAlgebra trait boundary for future geometric algebra
• Add Hopfield-Fenchel-Young energy-based associative retrieval
• Multi-scale encoding, morphological decomposer, fault-tolerant federation
• Improve SOTA for analogy, planner, and concept synthesis
• Add cost/total_cost fields to planner N-API types
• Evidence-weighted multi-hop confidence, ranked results
• Expose cleanup_vector N-API method for standalone Hopfield denoising
• Add agency layer (GoalStore, Planner, QuestionGenerator, SelfModifier)
• Add cognition layer, distributional refiner, CI coverage
• N-API bindings, persistence, load_from_log, compact for meaning memory (G6)
• Integrate meaning memory into HmsCore, wire write/delete paths (G5)
• Fuzzy_structural_query + multi_hop_query pipelines (G4)
• AtomMemory, CompositeMemory, TripleStore, RuleStore, Decomposer (G3)
• Shared utilities + IndexedMemory with Hopfield attractor (G1-G2)
• Rename to holographic-memory, fix npm logo, update all URLs
• Graph engine with multi-hop traversal, inference, temporal, federated queries
• Complete security integration, DP wiring, docs, LF readiness
• Add security features behind feature flag for LF Decentralized Trust readiness
• Migrate to writerslogic org, add npm publish pipeline
• Eliminate remaining gaps, expose train_nsg/train_ivf via N-API
• Expose diffusion config through HmsConfig and N-API
• Configurable thresholds, N-API config constructor, status getters
• Add query_sequence, elevate test coverage for weak features
• Add multi-shard support with auto-sharding
• Initial release of Holographic Memory System (HMS)

Changed¶

• Centralize SDK HTTP client and add CrewAI + Embedchain platforms
• Consolidate LangChain, LlamaIndex, and Haystack adapters into unified holographic-sdk
• Remove dead self-inverse resonator; move ResonatorConfig to phase_resonator
• Extract shared wire codec (write/read_lp_str, write/read_deltas)
• Extract meaning_ctx() helper shared by structural_query and multi_hop
• Gate experimental VSA modules behind experimental feature
• Migrate HMS agent identity from did.cose to CAWG ICA
• Deduplicate RNG pruning, unify patterns, add doc comments
• Eliminate dead code, add delete/compact with persistence
• Remove redb dependency, use in-memory FxHashMap with arena log persistence

Documentation¶

• State the AGPL-3.0-or-later license in the security policy
• Update changelog [skip ci]
• Record algorithm findings and private-search threat models
• Rewrite README to highlight Zero-Trust Enterprise Security and Python SDK
• State the testing policy explicitly in CONTRIBUTING.md (#66)
• Sync changelog (#54)
• Align guidance and track deferred production work (#52)
• Update changelog [skip ci]
• Update changelog [skip ci]
• Standardize badges (add OpenSSF Scorecard; consistent order/format)
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• §29 confirmed on unseen seeds — 4-bit matches float, 8-16x footprint win; log floor-exploration findings
• Wire deterministic-resonator capability into RESEARCH.md (discoverability + correct citation)
• Overnight closeout — queue exhausted, 337 tests green, flag pre-existing §21 bin lint
• Correct resonator-factorize repro command — bin is std-only, no experimental feature
• Reproducibility header for resonator-factorize (fixed seeds, correct repro command)
• DETERMINISTIC-RESONATOR.md — §20 validation write-up from in-loop 24-seed run
• Block codes also floor (§24); capacity gap is retrieval-vs-information, open not impossible
• Retract 'info-theoretically impossible' framing; 0.1D Plate is a readout artifact, closing gap to ceiling is open (§22)
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Record why presence gate uses an absolute margin, not SNR
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Pre-register non-self-inverse binding + nonlinear readout experiment
• Update changelog [skip ci]
• Update changelog [skip ci]
• Replace logo with continuous-rotation animation
• Update changelog [skip ci]
• Update changelog [skip ci]
• Restructure README with collapsible sections
• Update changelog [skip ci]
• Update changelog [skip ci]
• Rewrite README — fix logo for npmjs, restructure with quick start first, improve clarity
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Update changelog [skip ci]
• Add agent-provenance stack cross-reference to README
• Update changelog [skip ci]
• Update all documentation for v0.5.0 accuracy
• Add meaning memory architecture, structural queries, attractor cleanup
• Add meaning memory architecture, structural queries, attractor cleanup

Fixed¶

• Remove unimplemented security claims and retired cloud code
• Add missing experimental feature gate to zero_trust_demo
• Add missing experimental feature gate to zero_trust_demo
• Resolve CodeQL alerts and README formatting
• Resolve clippy::chunks_exact_to_as_chunks lint regression on main (#61)
• Resolve workflow code scanning alerts (#53)
• Give relations a distinct magic byte (0xFA), accept legacy 0xFE on read
• Clippy needless_range_loop in resonator-bundle (§21 bin) — enumerate rewrite, outcome-neutral
• Close remaining provenance trust-anchor gaps (cawg, vc-only records)
• Use sort_by_key to satisfy clippy 1.96 unnecessary_sort_by
• Repair --all-features build (port SCITT to ureq 3 API, drop dead CliffordVec bench, gate provenance example)
• Allow CDLA-Permissive-2.0 license from webpki-roots via ureq
• Replace bare unwrap() with expect() or total_cmp in production code
• Remove dead CliffordVec code and fix clippy warnings
• Perf and safety improvements from audit
• 5 bugs from audit (deadlock, double-decay, over-count, double-start, stale-used)
• Cargo fmt
• Use sort_by_key for clippy on newer Rust
• Cargo fmt formatting
• Benchmark crate rename, cargo fmt
• Resolve all clippy warnings for CI
• Quality pass on G1-G4 code
• Use absolute URL for logo on npmjs, add project.yaml for orchestrator
• Resolve 15 audit findings (3 critical, 8 high, 4 medium)
• Crash-safety, error propagation, and compaction correctness
• Propagate errors from shard insert/remove, guard ShardManager invariant
• Use .clamp() instead of .max().min() to satisfy clippy

Performance¶

• Optimize Zero-Trust FHE-Lite key generation and eliminate string cloning in hot loops
• Use posting-list candidate generation for synthesizeConcepts (>500 vectors)

Security¶

• Remediate repository alerts and dependencies (#38)
• Harden signing-key file perms, zeroize copies, bound JUMBF recursion

Assets¶

• Static transparent logo from first frame, text removed
• Simplify logo — strip floating decorative specks, keep the connected graph mark
• Replace logo with self-contrasting node-graph mark (animated gif + static svg), drop garbled-text variants

Deps¶

• Land 10 Dependabot major bumps with API fixes (#26)

Example¶

• Story_memory -- HMS reasoning memory serving a writer (scrivener-mcp)

Research¶

• Fractional-power encoding gives the phasor memory continuous 'near'
• Phase memory retrieves under load where sparse collapses
• Rotation binding gives relation algebra the current substrate lacks
• Path-plasticity -- holographic generalization a cache can't do
• Living-connection-graph slice -- plasticity beats saturation, verifiably

• 2 hardening RETRACTS the sparse-wins claim (AUC artifact)¶

• 2 strong baselines HRR + MAP vs sparse permutation¶

• 2 involution control disconfirms the self-inverse framing¶

• 2 density-matched control rules out the confound¶

• 2 step-1 binding discriminator harness + result¶

Style¶

• Apply rustfmt to experiment binaries