OpenBMB open-sources T2Mem for test-time memory construction
OpenBMB released the T2Mem repository on GitHub, providing experiment code for test-time memory construction in LLMs. The code is MIT-licensed and requires Python 3.10+.
China context
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can use the MIT-licensed code to implement test-time memory construction in their own LLM applications, potentially reducing memory overhead.
- For investors
- The release signals OpenBMB's continued investment in memory-efficient LLM techniques, which could lower costs for long-context applications.
OpenBMB published the T2Mem repository, which contains experiment code for the T2Mem method. T2Mem enables LLMs to reason over long interaction histories by deferring memory construction to test time. It extracts lightweight episode-level metadata offline, then adaptively retrieves and organizes metadata and raw interactions to progressively build query-specific memory as reasoning unfolds. The repository includes setup instructions, reproduction guides, and links to the LoCoMo and LongMemEval benchmarks. The code is MIT-licensed and requires Python 3.10 or newer.
T2Mem defers memory construction to test time, extracting lightweight episode-level metadata offline and then adaptively retrieving and organizing metadata and raw interactions to build query-specific memory as reasoning unfolds. The repository provides experiment code and reproduction guides for the LoCoMo and LongMemEval benchmarks.
Developers working with long-context LLMs gain a new open-source method that reduces memory construction overhead while improving performance on long interaction histories, potentially lowering the cost of building memory-augmented applications.
The MIT license allows commercial use, and the method's reduced memory construction overhead could lower infrastructure costs for applications requiring long-term memory in LLMs.
The next verifiable signal is whether the T2Mem paper is published on arXiv and whether independent evaluations on LoCoMo and LongMemEval confirm the reported performance gains.