memnest v0.3.1

Changes

• Replay explicitly selected transcript history with watch --backfill --once --path /original/absolute/path, without rewinding live watcher cursors. Add --dry-run to count eligible chunks without writing. Eligible does not mean missing; original-path identity and existing deletion rules still apply.
• Stop padding mixed Korean/Latin named searches with candidates that contain none of the explicit Latin identifiers. Semantic fallback remains when no named candidate exists. This is not a general relevance or answer-quality guarantee.
• Add an optional following_id hint to at most one retrieved user capture. Read that source with memory_get instead of fetching an entire neighborhood. The hint follows stored capture order, not a verified reply relationship or a guarantee of a final answer. Existing search result count and order are unchanged.
• Report that past-history completeness has not been checked.

No database reset, automatic historical import, new dependency, or additional generative model call is required. Existing stores and deletion intent are preserved.

Downloads

• Linux x86_64 and aarch64 core archives, each with a SHA-256 checksum.
• pi-memnest-0.3.1.tgz and its SHA-256 checksum. Unpack to a persistent directory, then use pi install /absolute/path/to/package to load the bundled extension. Remove an existing Memnest package declaration before registering a different source to avoid duplicate tools.

pi-memnest 0.3.1 is now available on npm. With the core running, install it with pi install npm:pi-memnest@0.3.1. This extension includes the previously unpublished 0.3.0 changes, including default-off automatic context and paged source reads. The npm package contains the same executable bundle as the GitHub archive, with updated installation documentation; the archives have separate checksums.

macOS remains source-only and is not a validated installation target. No Windows binary archive is published.

Validation and limits

The exact release source passed 120 Rust library tests, 5 binary tests, formatting, Clippy, license checks, pi bundle and smoke checks, and real stdio MCP tests. Local core and extension archives were unpacked and executed. GitHub's release workflow separately builds and checks the Linux archives; aarch64 receives archive-layout checks rather than native runtime validation.

Across 15 selected historical questions, 59 search results and their order were preserved in a contemporaneous comparison. Two following-source hints led to the expected stored explanations. Search responses grew by 264 characters in total; narrower source reads reduced returned text in those two paths. This was a retrieval-path test, not a generative answer-quality test or a measurement of billed-token, latency, or total conversation savings.

See the core changelog, pi changelog, and following-source limits.

Full changes: v0.3.0...v0.3.1