Skip to content

Improve memory grounding, retrieval, and evaluation fidelity - #3212

Merged
Hillary Mutisya (hillary-mutisya) merged 5 commits into
mainfrom
dev/hillarym/memory-quality
Oct 24, 2026
Merged

Hillary Mutisya (hillary-mutisya) merged 5 commits into
mainfrom
dev/hillarym/memory-quality

Conversation

@hillary-mutisya

Copy link
Copy Markdown
Collaborator

Summary

  • Preserve source-owned temporal context and provenance through ingestion, retrieval, and answering. Reject invalid optional date annotations without failing valid imports.
  • Improve default evidence retrieval and packing with complete query-term coverage, source-scoped fallback, score-first selection, and overlapping-snippet removal.
  • Ground source synopses in host-owned evidence IDs and add ingestion, embedding, and storage profiling.
  • Align LongMemEval with the product memory factory and default answer configuration, with resumable runs, frozen checkpoints, and precise diagnostics.

Validation

  • Full workspace build passes.
  • 300 focused product tests and 196 offline qualification/harness tests pass.
  • Formatting, lint, complexity, dependency-cycle, and test-debt gates pass against main.
  • Earlier inspected ten-case LongMemEval regression: 476/476 sessions imported, 6/10 answers correct, zero ingestion failures. This was not rerun after merging main and is not a holdout or full-suite accuracy estimate.

Compatibility

Existing indexes remain readable without automatic migration. Reindex existing corpora to regenerate derived temporal evidence under pipeline version 2.

Preserve source-owned temporal evidence, improve scoped retrieval and evidence packing, and ground source synopses in host-owned evidence IDs.

Align LongMemEval with product memory defaults and add reproducible qualification, diagnostics, and ingestion profiling.

Squashed from dev/hillarym/memory22 at ab1c437; original branch retained.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace overlapping whitespace matching with a single-line capture and separate leading-whitespace normalization. Preserve source date spelling and scope offsets without borrowing dates from subsequent lines.

Add LF/CRLF, spelling, scope-boundary, and 100,000-tab regression coverage; document single-line date headers.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Preserve memory-quality checks in the split CI script phase and retain both memory testing documentation sections.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Run entrypoint guard checks independently from synchronous scorecard tests and assert their expected validation errors. Document why the offline suite remains suitable for blocking CI.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@hillary-mutisya
Hillary Mutisya (hillary-mutisya) merged commit 32417f0 into main Oct 24, 2026
30 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants