vshulcz/deja-vu is a Go binary that indexes on-disk session logs from 42 coding agents, including Claude Code, Codex, and Cursor. The latest release is v0.22.1. The provided FACTS list no specific items changed in this version, so the release changes cannot be confirmed. According to the project description and README summary, it keeps the logs local and searches them without a large language model, and delivers results to agents through an MCP server and event hooks. The license is MIT.
The purpose of this tool is to let agents find again what they need for their work in the session logs they have already left behind. Since multiple agents can read the same index, it is a way to search the logs of each tool in one place. The README states that it recorded R@5 of 97.2% on the LongMemEval-S evaluation, and claims this is higher than MemPalace at 96.6% and agentmemory at 95.2% on the same material. Indexing speed is reported as 19,195 sessions processed in 17.6 seconds, while the seven comparison tools took 72 seconds. Token savings are described as half of agentmemory on already-resolved tasks. These figures are evaluation results published by the project, not guarantees of performance in every real user environment.
When interpreting the evaluation results, the scope of the evidence matters. The comparison baselines, the corpus, and the scoring rules are all presented in the maintainer's repository, and the validation used a synthetic corpus rather than real user logs. It is therefore hard to conclude from the published numbers alone that the same search quality will hold on actual session logs. The README also reports a problem where only 77% of decisions and 0.2% of commands remain in the summary after compaction. It focuses on searching session logs, but it is hard to consider it a tool that preserves even the content lost during compaction.