Repository navigation
Releases
v0.3.4
Fixed
- Allow
import --global/import -gwith exported JSONL files and directory
imports without requiring--agent. Imported documents retain their metadata
and vectors and are appended to the global collection. - Download Jina's external ONNX weights alongside its model graph and detect
incomplete model installations during setup.
Changed
- Make the Jina v5 nano retrieval benchmark settings the fresh-install defaults:
768-dimensional embeddings, last-token pooling, vector-BM25 fusion with a
lexical weight of 0.10, 300 candidates, and cross-encoder reranking disabled. - Align the example configuration, development model downloads, and setup
documentation with Jina. - Automatically update the Homebrew formula after release archives are verified
and published.
Existing configuration files continue to override defaults. To switch an Arctic
embedding database to Jina, export with --no-embeddings and import into a fresh
database to regenerate compatible vectors.
v0.3.3
Fixed
- Publish a complete
checksums.txtcovering all five macOS, Linux, and Windows
archives. This fixes Linux installer failures caused by missing checksums. - Keep releases in draft until every platform build, archive check, and combined
checksum manifest succeeds, solatestnever points at a partial release.
Changed
- Build with Go 1.26.8 instead of 1.26.0.
- Replace the deprecated Windows MINGW64 toolchain with UCRT64 and explicitly
statically link native compiler support libraries. Windows system DLLs and
the separately provisioned ONNX Runtime remain runtime dependencies. - Use
-trimpathand-s -wconsistently for published executables. - Run Windows Go tests in CI, with portable home/data-directory and invalid-path
fixtures, in addition to the extracted-binary startup and SQLite smoke test. - Expand Windows troubleshooting guidance for Defender ASR reputation blocks.
These build and packaging changes are not a confirmed fix for Defender ASR rule
01443614-cd74-433a-b99e-2ecdc07bfc25. Windows releases remain unsigned unless
maintainers configure signing; acceptance on managed devices requires a retry
under the device's existing policy. version and --help do not require ONNX
Runtime or models.
v0.3.2 — Windows startup diagnostics and release validation
Mnemoteca v0.3.2 improves Windows release validation and troubleshooting on restricted devices.
Windows retry and diagnostics
- Download
mnemoteca_0.3.2_windows_amd64.zip, extract it, and runmnemoteca.exe versionfrom PowerShell. - Follow the Windows startup and offline provisioning guide. A copy is also included at
docs/windows.mdin the ZIP. - The
.zip.sha256sidecar contains SHA-256 values for both the ZIP and extracted executable.
This release's Windows executable is unsigned. Release signing support has been added, but the signing service is not configured. A restricted device may still block startup. The guide explains how to collect Code Integrity/AppLocker evidence and obtain approval through your organization's application-control policy without disabling protection. Local inference does not exempt the executable or ONNX Runtime DLL from that policy.
Changes
- Add optional Azure Artifact Signing through OIDC, with signature and timestamp verification before packaging.
- Validate Windows startup and SQLite initialization in an isolated directory without compiler paths or models; validate the extracted release ZIP before upload.
- Include Windows trust-policy diagnostics and instructions for provisioning models on machines without outbound access.
- Include installer improvements since v0.3.1: verify release archive checksums, simplify migration prompts, and keep compatibility-symlink reinstalls quiet.
Full comparison: v0.3.1...v0.3.2
v0.3.1
What's New
Migration from Mnemosyne to Mnemoteca now preserves your embedding settings, including 768-dimensional legacy models, so existing memory exports import correctly. Public installation also avoids GitHub API rate limits.
Detailed Changelog
New Features
- Migrate legacy Mnemosyne configuration settings to Mnemoteca during guided installation.
- Import existing JSONL exports using the dimensions of their stored embeddings.
Bug Fixes and Improvements
- Rewrite the migrated database path to the new Mnemoteca data location while preserving legacy model and search settings.
- Allow public installers to resolve releases without GitHub REST API authentication or rate-limit failures.
Breaking Changes
None.
v0.3.0
What’s New
- Mnemoteca replaces Mnemosyne. The CLI, installer, release artifacts, documentation, and configuration paths now use the Mnemoteca name. The release includes a guided migration path for existing Mnemosyne data.
Changelog
Others
- e10338f Fix integration test document ID parsing
- a354c32 Harden installer test PTY harness
- 5477120 Make installer tty detection portable
- e1e3d57 Propagate installer harness diagnostics
- e41fcd1 Rename CLI to Mnemoteca
- faf744d Report installer harness failures
- 7bf7b7e Show installer exit status in harness
- b7a7bc5 Use supported no-terminal harness session
v0.2.7
New Features
- Lazy global collection initialization: The
globalcollection is now silently created on first explicit--global/-guse, removing the need to runmnemosyne init -gas a prerequisite. Named and project collections remain strict (must be initialized first).
Bug Fixes and Improvements
- Refactored 7 command paths (add, search, list, tags, update, forget, export) to use a shared
getSelectedCollectionhelper, reducing duplication - Fixed misleading error message:
mnemosyne init --name globalwas previously suggested but rejected by the reserved-name validator
v0.2.6
New Features
- Native Linux release binaries: Publish
linux/amd64andlinux/arm64archives built on native GitHub Ubuntu 22.04 runners. Replaces the old Zig cross-compilation approach with native CGO builds targeting an Ubuntu 22.04-era glibc baseline. ONNX Runtime and models remain installed separately viamnemosyne setup.
Bug Fixes and Improvements
- Remove obsolete
.goreleaser.default.yamland stale commented Linux/Zig build block - Rename
download-all-libstodownload-darwin-libs, strip unused Linux library downloads - Remove
zigpreflight from localtask release - Add ARM64 compile-only CI job to catch CGO regressions before tag push
- Update README release documentation for tag-push workflow and native Linux strategy
- Mark Linux releases as complete in ROADMAP
Breaking Changes
- None
v0.2.5
New Features
mnemosyne updatecommand — Strict in-place document editing by ID within a collection. Supports positional text,--file,--stdin, additive tag updates (--tag), tag replacement (--replace-tags), and full content+vector replacement with transactional rollback on error.
Bug Fixes and Improvements
- CI verification task — Automated
lint && testtask for pull requests. - Improved README — Surface integrations under Installation, document Claude Code setup with npx skills, update generic agent memory instructions, refresh Quick Start examples.
Breaking Changes
None.
v0.2.4
New Features
- Claude Code memory import — Import existing Claude Code memory files into a Mnemosyne collection with
mnemosyne import --agent claude. Supports dry-run preview (--dry-run), user-level memories (--include-user), custom paths (--path), and global collection (--global). Non-destructive: only reads, never modifies source files.
Bug Fixes and Improvements
(None in this release)
Breaking Changes
(None in this release)
v0.2.3
Changelog
v0.2.3 (2026-06-23)
New Features
- YAML config loading and JSON search output — Load config from
~/.config/mnemosyne/config.yamlwith env override support. Add-f jsonoutput format for search. Support customonnx_filein reranker config andtask_idONNX input for embedding models. (c9b75a3) - Vector-first BM25 fusion search — Default search now retrieves vector candidates first, then applies in-memory BM25 reranking with blended scores. Add
--fusion,--bm25-weight,--fts-only,--vector-onlyCLI flags. (dc960dc)