mnemosyne-engine v1.1.0

Document ingestion, partitioned index, and schema ingestion.

See CHANGELOG.md for full details.

Highlights

• PDF, DOCX, CSV, plaintext extraction (zero-dep core, optional pypdf)
• Isolated document partition with BM25 + TF-IDF retrieval
• Schema ingestion: SQL DDL, JSON/YAML, SQLite introspection
• CLI --docs and --all flags for partition-aware search
• One-time upgrade hint for existing users
• 446 tests, 0 regressions

Upgrade

Run mnemosyne ingest --full after upgrading to populate document partitions.

mnemosyne-ollama v0.1.1

Partition visibility and version fix.

Improvements

• Verbose mode (-v) shows code/docs chunk counts and file paths per partition
• --version flag reads from package version instead of hardcoded string

mnemosyne-mcp v0.2.0

Federated search, new tools, and bug fixes.

New Tools

• search_docs -- document-only search
• schema_ingest -- DDL/JSON/SQLite schema ingestion
• schema_stats -- schema index statistics

Fixes

• Per-partition truncation (docs no longer silently dropped)
• Stats handler crash fix
• Document TF-IDF vocabulary wiring

v1.0.5

Fix PyPI README rendering (broken image, dead links). Replace Mermaid diagrams with PNGs in REFERENCE.md. Add trademark notices.

mnemosyne-ollama v0.1.0

Lightweight MCP host for Ollama. Zero-config local code search with tool-calling models. pip install mnemosyne-ollama

v1.0.4 — Verified publishing + self-ingestion fix

What's New

Verified PyPI Publishing (v1.0.4)

• Added OIDC trusted publisher workflow for verified releases on PyPI
• All GitHub Actions SHA-pinned to prevent supply chain attacks
• Environment-gated deployment with required reviewer approval
• Signed SLSA artifact attestation on every release
• Tag-to-version verification prevents stale builds

Bug Fixes & Improvements (v1.0.2)

• Fixed self-ingestion bug where mnemosyne could index its own database files
• Fixed benchmark root path resolution
• README overhaul — clearer install, usage, and architecture docs

Security

• Zero secrets/tokens in CI — pure OIDC authentication
• Least-privilege permissions on every workflow job
• Build and publish are isolated jobs (build env cannot publish)

Full Changelog: v1.0.2...v1.0.4

v1.0.3 — Verified publishing + self-ingestion fix

What's New

Verified PyPI Publishing (v1.0.3)

• Added OIDC trusted publisher workflow for verified releases on PyPI
• All GitHub Actions SHA-pinned to prevent supply chain attacks
• Environment-gated deployment with required reviewer approval
• Signed SLSA artifact attestation on every release
• Tag-to-version verification prevents stale builds

Bug Fixes & Improvements (v1.0.2)

• Fixed self-ingestion bug where mnemosyne could index its own database files
• Fixed benchmark root path resolution
• README overhaul — clearer install, usage, and architecture docs

Security

• Zero secrets/tokens in CI — pure OIDC authentication
• Least-privilege permissions on every workflow job
• Build and publish are isolated jobs (build env cannot publish)

Full Changelog: v1.0.2...v1.0.3

v1.0.0 — Dense Embeddings, Retrieval Quality Upgrades, Production Release

What's New

Dense Embedding Backend

• Optional 23MB ONNX model (all-MiniLM-L6-v2-code-search-512) as 6th retrieval signal
• Opt-in via config, downloads on first use, runs 100% locally, zero data egress

Retrieval Quality

• Porter stemmer in TF-IDF — aligns with BM25/FTS5 porter stemming
• Code-aware stopwords — preserves get, set, for, is, has in code identifiers
• Two-pass soft file filter with chunk-qualification
• Symbol match file promotion — guaranteed filter slots
• FTS5 escape fix — commas/periods/hyphens no longer break BM25
• Type-aware symbol boost — 3x all, 4x for class definitions with PascalCase
• Decorator inclusion in function/class chunks

Benchmarks (Claude Opus 4.6, real codebases)

| Production repo (829 files) | Open-source (100 files) |
|---|---|---|
| Token cost | 12K vs 45K (73% savings) | Parity |
| Correct file | 100% (20/20) | 100% |
| Answer quality | Equivalent to baseline | Equivalent |

Install

pip install mnemosyne-engine

Full changelog: CHANGELOG.md

Mnemosyne v0.3.0 — Initial Open Source Release

Zero-dependency Python 3.11+ context compression and retrieval engine for LLM agents. Reduces context waste by 40-70% with sub-100ms queries.

Highlights:

• Hybrid BM25 + TF-IDF retrieval with 7-signal RRF fusion
• AST chunking (Python), regex+brace chunking (Go, C#, Rust, Java, JS/TS)
• 4-stage compression with control flow preservation
• ARC cache, Bloom filter, staleness detection, daemon mode
• 293 tests, zero runtime dependencies

Get started:
pip install -e .
mnemosyne init
mnemosyne ingest
mnemosyne query "how does auth work"

Dual-licensed: AGPL-3.0 (open source) + Commercial.