doronp/gitmemory is a memory tool that stores agent conversations verbatim in a local Git repository and finds the needed location by byte offset. The repository's latest release is v0.1.0. What this version offers at its core is a way to look up the supporting evidence in the original conversation even after context compression.
Compressing a long conversation can push some content out of the context. gitmemory stores the conversation bytes, verifies with a hash whether the segments join without gaps, and then searches. According to the README, among470 LongMemEval instances, on the cases where evidence exists after the compression boundary turn recall was 0.7456, while the live context window under the same conditions was 0.0000. LLM-free retrieval using a local cross-encoder ranked first on LongMemEval-S with session R@5 of 99.2 and R@10 of 99.8, and ranked second on all-evidence@10 and LoCoMo. There are no LLM calls or network requests at runtime, and there are zero core dependencies. It requires Python 3.13 or later and git, and is licensed under Apache-2.0.
These numbers are results cited by the README, and there is nothing in FACTS that independently confirms benchmark performance. The README states that it also publishes the runs that failed in the last benchmark. So while the byte preservation and the search method are concrete, it is more reasonable to keep any performance judgment within the bounds of these published self-reported results.