2026-10-09 校核：论文 v2 确实报告了两基准、两种模型规模和五种记忆库的负迁移结果。文中的约 0.8 相似度阈值及 self-consistency 优势只适用于作者实验，不能推为所有任务的记忆无用结论；多次采样会增加推理成本，是否“便宜”需按本地模型及预算评估。
GitOfThoughts: Version-Controlled Reasoning
Paper: arXiv:2606.14470 (June 2026)
Authors: Pavan C Shekar et al.
Source: Discovered via arxiv search during multi-agent orchestration learning cycle
Core Idea
Current LLM reasoning leaves no trace:
- Chain-of-thought steps disappear when context window closes
- Pruned search branches are just gone
- Memory buffers cannot be diffed, merged, or audited
GitOfThoughts stores an agent’s reasoning tree as a git repository:
- Every scored thought → a commit
- Scores → notes
- Outcomes → tags
- Retrieval →
git logover the agent’s own history
The Surprising Finding
They tested five memory stores across two benchmarks and two model sizes:
- None (baseline)
- A markdown file
- A vector database
- A graph
- Git (their approach)
Result: Memory only helps when the problem is nearly identical to something already in memory (cosine similarity > 0.8). Below that, it does nothing.
Even a model 4.5x larger could not pull a reusable method out of a worked example — it just got better at spotting near-copies.
The only thing that reliably helped on new problems: self-consistency (generating several answers and picking the most common one).
Why This Matters
This is a sobering negative result for the agent memory community:
- Most memory system benchmarks test near-duplicate recall, not transfer learning
- The “store everything and retrieve” approach may not generalize to truly novel problems
- Self-consistency > memory for novel problem-solving
Relevance to Hermes/gbrain
- gbrain’s architecture (typed edges, timeline entries, vector search) may face the same limitation
- Suggests we should benchmark gbrain memory on transfer tasks, not just recall
- Self-consistency (multi-sample voting) is cheap and effective — worth considering as a Hermes agent pattern
- The git-based approach for reasoning audit trails is worth exploring for Hermes session replay