TMCRA is now positioned as an Agent Memory Engine for scope-isolated, source-traceable long-term memory.
Highlights:
- Publishes the LongMemEval S500 scorecard: 411/500 (82.2%) with six task-level results.
- Adds English and Chinese architecture and reproduction documentation.
- Adds the maintained, security-hardened LongMemEval pipeline and benchmark-bound inference checkpoints through Git LFS.
- Uses the Apache License 2.0.
- Credits Yu Haoxin as creator and lead developer, with OpenAI Codex credited for development and reproducibility engineering assistance.
See the repository README and CHANGELOG for the complete release scope.