v0.2.0

First cleaned public release of Mstar — Every Task Deserves Its Own Memory Harness (arXiv:2604.11811).

The code that produces the paper's results is unchanged. This release brings the repository in line with the published paper and removes development scaffolding.

Documentation

• Promoted the full guide to the root README.md (assets moved to assets/).
• Aligned all docs with the paper: correct title, the four paper benchmarks (LoCoMo, ALFWorld, HealthBench, PRBench), and the paper⇄code term bridge (memory program / memory harness == KBProgram).
• Documented reproduction via scripts/run_experiments.sh (Azure gpt-5.4-mini / gpt-5.3-codex), with a clear warning that the DeepSeek CLI defaults do not reproduce published numbers.

Cleanup

• Removed unused DVC integration, the internal AGENTS.md, and development-only scripts.
• Removed Data/STATE-Bench/ (an unrelated vendored dataset) and its wiring.
• Regenerated uv.lock.

v0.1.1

What's New

• FastEmbed local fallback — --embedding-model local 跳过 API，使用本地 ONNX 模型；API 失败时自动降级
• --output-dir 自动断点续跑 — 每个运行有唯一输出目录，中断后重新运行自动恢复
• 实验扩展至 7 个数据集 — LoCoMo, ALFWorld (unseen/seen), HealthBench×2, PRBench×2
• 环境变量覆盖模型 — TASK_MODEL / REFLECT_MODEL / TOOLKIT_MODEL / EMBEDDING_MODEL
• 实验运行指南 — scripts/README.md，中文，含环境配置、模型切换、断点续跑说明

Bug Fixes

• Fix ALFWorld seen split mapping (valid_seen → valid_train)
• Fix embedding API crash with proper FastEmbed fallback
• Fix PRBench hard item deduplication
• Unify Scorer/ValScorer protocols

v1.0.0

Initial release of Engram: Memory Is a Program — Evolving Agent Memory Through Executable Code Search.