v0.5.0

v0.5.0 — Evaluation Foundation

Added

• Candidate-neutral recall-v3 evaluation with strict corpus contracts, digest commitments, isolated lexical-baseline runs, and stable reports.
• Offline manifest-driven exact-vector candidate validation for semantic-only, lexical-reranking, and union experiments.
• Operational, task-utility, privacy-safe trace, deterministic workload, and competitor-evidence validators.

Important scope

This release ships the evaluation foundation only. The included v0.5 fixtures are synthetic harness contracts, not evidence that semantic or hybrid recall should ship. Semantic retrieval remains undecided and normal product operation continues to use lexical recall without model installation, network access, or an embedding index.

Any future semantic profile remains opt-in and requires the independently governed evaluation process described in the roadmap.

Memzoi v0.4.0

Memzoi v0.4.0 delivers a measured, review-first evidence-backed capture loop for file-native project memory.

Highlights

• Adds file-native recall v2 evaluation suites covering search, precheck, context, write gates, lifecycle/scope leakage, citations, provenance, latency, and token use.
• Adds deterministic capture from explicit Markdown, agent instruction files, ADRs, supplied Git diffs, and immutable Git ranges, with exact evidence and generated-content exclusion.
• Keeps capture planning mutation-free and routes reviewed candidates through existing private/proposal/canonical boundaries, with secret, PII, and transcript blocking.
• Adds capture v1 evaluation gates for candidate quality, exact evidence, prohibited output, mutation safety, latency, and human review burden.
• Adds the planning-only MCP plan_capture_v1 tool with bounded, structured, redacted behavior.
• Records the accepted evidence-backed capture/extractor contract in RFC 0001.

Evaluation baseline

All 23 capture cases pass: candidate precision and recall are 25/25, exact evidence is 33/33, all 4 forbidden opportunities are suppressed, and every capture hard gate is zero.

Known limitation

CLI and MCP capture file operations fail closed on Windows because v0.4.0 requires Unix handle-relative, no-symlink file access. Windows release binaries remain available for the rest of the CLI and MCP surface.

Full changelog: v0.3.1...v0.4.0

Memzoi v0.3.1

Memzoi v0.3.1 completes the Trust Contract across recall eligibility and reviewed proposal lifecycles.

Added

• Inspect expiry decisions through CLI memzoi expiry and MCP inspect_memory_expiry. (#40)
• Resolve reviewed create, reject, supersede, and tombstone proposal packets into Git-readable outcomes. (#43, #46)
• Track approving proposal lineage separately from original evidence provenance. (#42)

Changed

• Expired records are excluded consistently from normal search, context, handoff, precheck, runtime reads, and generated exports using the inclusive now >= expires_at boundary. Canonical files remain unchanged. (#40)
• Every canonical apply route requires an explicit repo-safe sensitivity. Omitted or legacy values become unknown, and blocked output remains structured and redacted. (#41)
• Evidence source and source_ref values survive apply, rebuild, recall, and export independently of proposal_id. (#42)

Fixed

• Path-bound governance memory is recalled without requiring lexical overlap, while preserving deterministic ranking and scope filtering. (#39)
• File-backed apply and reject operations resolve their packets and synchronize relational and full-text recall before returning. Reported failures roll back affected state, and replay repairs derived drift. (#43)
• Supersede and tombstone packets validate one active, same-scope, non-stale target before mutation. (#46)
• OKF projection and import content hashes match the persisted trimmed body.

See CHANGELOG.md for the complete release history.

Memzoi v0.3.0

Added

• Two-plane Git/runtime memory policy with explicit destination, write-route, review, and provenance metadata.
• Local-only runtime records, session checkpoints, session-end promotion, layered context ranking, and handoff packs.
• Review-first classified imports with guarded repo proposal, local runtime, and session checkpoint writes.
• Deterministic Codex, Claude, and MCP integration profiles generated from the canonical memory policy.
• JSONL event export and expanded proposal-file review/apply workflows.

Changed

• Recall provenance is now serialized as the storage plane (git or runtime), while destination remains the routing value (repo, local, or session).
• memzoi proposal-files apply reports the required rebuild, and memzoi doctor detects runtime repo-index drift.
• Release builds support non-uploading cross-platform dry runs and validate version, tag, and docs metadata before publishing.

Fixed

• Downloaded checksum sidecars use archive basenames and portable LF-only manifests.
• Pull requests targeting stacked feature branches now run Rust CI.

See CHANGELOG.md for the complete release history.

Memzoi v0.2.0

What's Changed

• add memzoi update command by @Zokiio in #1

New Contributors

• @Zokiio made their first contribution in #1

Full Changelog: v0.1.0...v0.2.0

Memzoi v0.1.0

Memzoi v0.1.0 is the first public baseline for the file-native AI memory experiment.

Install on Mac or Linux:

curl -fsSL https://raw.githubusercontent.com/Zokiio/Memzoi/main/scripts/install.sh | sh

Install on Windows:

powershell -ExecutionPolicy Bypass -c "irm https://raw.githubusercontent.com/Zokiio/Memzoi/main/scripts/install.ps1 | iex"

Highlights:

• Release binaries are attached for macOS arm64, macOS x64, Linux x64, and Windows x64.
• Install scripts download the latest release binary by default and verify SHA-256 checksums.
• Cargo is only required for source installs, for example make install or MEMZOI_REF=main.
• Docusaurus docs are versioned for 0.1.0 and keep next docs available separately.
• crates.io publishing is prepared but not yet published.