deja-vu: A DSH plugin That Shares One Memory Across Coding Agents for DeepSeek Harness
vshulcz/deja-vu
deja-vu indexes the session history your coding agents already wrote to disk, giving DeepSeek Harness and 34 other agents one shared memory with millisecond recall — no LLM, no embeddings, one local Go binary.
deja-vu is a DSH plugin that turns the coding session history already sitting on your disk into one shared memory. It targets the same issue being re-debugged in a different agent: a fix found in one agent comes back in any of them, including months of sessions from before you installed it. A single local Go binary handles indexing and millisecond lookups with no LLM and no embeddings, and keys and tokens are stripped as the index is built. Installation wires recall into every agent it detects and builds the first index in seconds.
How to Install
dsh plugin --profile web add dsh-deja- Category
- Tools & Capabilities
- Platform
- DSH Plugin
- Author
- vshulcz
- Distribution
- Plugin
deja-vu Key Features
deja-vu Repo Summary
What Does It Do?
Deja Vu is a DSH plugin for DeepSeek Harness that turns the session history already sitting on your disk into one shared memory across coding agents. It solves the problem of re-debugging the same issue in a different agent: a fix found in one agent comes back in any of them. Maintained by vshulcz under the MIT license, it is a single local Go binary with no LLM and no embeddings, so indexing and lookup stay on your machine.
Core Features
- Shared memory across agents: indexes the session history written by Claude Code, Codex, Cursor, Copilot CLI and more than thirty other coding agents, so any agent that asks gets the same result.
- No model, no embeddings: a local Go binary handles indexing and millisecond lookups over gigabytes of history, with no capture step and no vectorization.
- Automatic recall: history arrives at session start, on every prompt, before a file is edited or a command runs, and after one fails, so the agent does not have to search.
- Redaction at build time: keys and tokens are stripped while the index is built, so what reaches the model is safe to send.
- Verifiable retrieval quality: 85.3% hit@1 on LongMemEval-S and 69.7% retrieval hit@1 on LoCoMo, with both harnesses shipped so you can check the numbers on public datasets.
- Useful immediately: installation wires MCP recall into every agent it detects and builds the first index in seconds.
How to Use This Plugin?
After enabling it in DSH, start a new agent session and ask about something you worked on months ago, for example whether you have dealt with JWT refresh rotation before; it pulls the matching session from shared memory. With auto-recall on, the agent already knows what was solved in that project when the session opens. For search-only use, the binary alone covers indexing, search, context and blame, with JSON output; a bundled search skill teaches an agent the CLI contract when MCP is not wired. If you prefer agents to call memory on their own rather than starting each session with it, use the install mode that omits session-start recall.
How to Troubleshoot This Plugin?
On Windows the install script exits with unsupported OS because it is a shell script; follow the official instructions and use the Windows package manager or take the release archive and put the executable on your PATH. If you run the binary without wiring it into agents, deja doctor reports every MCP target as not-wired, which is that setup working as intended; complete the wiring as described in the official guide.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: vshulcz/deja-vu. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.