Changelog

All notable changes to Mnemosyne are documented in this file.

Format follows Keep a Changelog. This project uses Semantic Versioning.

1.0.0 - 2026-03-30

Dense embedding backend — optional 23MB ONNX model (all-MiniLM-L6-v2-code-search-512) as 6th retrieval signal via Reciprocal Rank Fusion. Bridges vocabulary gap when query terms don't match code identifiers. Opt-in via config, downloads on first use, runs 100% locally. No data egress.
Porter stemmer for TF-IDF tokenizer — aligns with BM25/FTS5 porter stemming.
Code-aware stopwords — get, set, for, is, has preserved in code identifier splitting instead of being stripped as English stopwords.
Two-pass soft file filter — files containing top-50 individual chunks survive at 0.7x penalty even if their aggregate file score is low.
Symbol match file promotion — files with strong symbol matches get guaranteed filter slots.
Decorator inclusion — function/class chunks now include decorator blocks.
Chunk enrichment — TF-IDF embeddings include file path and module docstring context for better semantic matching.

mnemosyne health command — reports index age, file/chunk/token counts, vocabulary size, stale files, FTS5 integrity, tokenizer hash, and daemon status. --json flag for programmatic monitoring.
--log-format global flag (text/json) for structured logging on all commands.
--log-level global flag (DEBUG/INFO/WARNING/ERROR) for log verbosity.
mnemosyne benchmark command for running retrieval-quality benchmarks against any corpus with JSON question sets.
httpx 0.28.1 benchmark regression gate (5 assertions, @pytest.mark.benchmark).
Retrieval metrics: hit_at_3, relevant_at_5, mrr_at_10 in BenchmarkSuite.
Symbol type-aware ranking — _symbol_search returns chunk_type; class definitions receive 4x boost when PascalCase/TitleCase query detected.
Hybrid file aggregation in _file_level_filter — max + 0.1*sum prevents files with many weak chunks from volume-dominating over precise matches.
Ingest directory walk — mnemosyne ingest src/ now expands directories instead of silently dropping them.
Path containment validation — ingest CLI rejects paths outside project root, resolves symlinks with os.path.realpath().

FTS5 escape regex — commas, periods, hyphens, and other punctuation silently caused BM25 to return 0 results for natural language queries.
_split_class_chunk now preserves parent chunk_type instead of hardcoding "block", enabling class-name boost for split class definitions.
Daemon socket permissions set to 0600 after bind (CVE-class prevention).
Daemon ingest path validation — rejects paths outside project root.
Six silent except:pass blocks replaced with proper WARNING/ERROR logging.

Ingest _scan_files() refactored into _scan_dir() shared helper for both full-scan and targeted path resolution.
_file_level_filter aggregation changed from pure sum to hybrid (max + 0.1*sum) — improves precision for single-file queries.

Initial public release on PyPI as mnemosyne-engine.