English | 简体中文 | 繁體中文 | 日本語 | 한국어 | Español | Português | Français | Deutsch | Русский | Türkçe | हिन्दी
Part of deja-vu (97.2% R@5 on LongMemEval-S · 13.6 s to index 19,195 sessions · half the tokens of agentmemory) · Website & live search · Browser demo · ★ Star deja-vu on GitHub
Cross-agent session memory for Claude Code, Codex, Cursor, Gemini CLI, OpenClaw, Hermes, and forty-seven other coding agents in total — injected at the Magpie gateway with zero LLM dependencies.
When you route multiple coding agents through a gateway, each client tracks only its own thread and loses context whenever you switch tools or start a fresh session. @vshulcz/magpie-deja recalls relevant prior sessions from your local deja BM25 index during Magpie gateway requests across OpenAI (chat, responses), Anthropic (anthropic), and Gemini (gemini) protocols.
magpie plugin add @vshulcz/magpie-deja@vshulcz/magpie-deja depends on @vshulcz/deja-vu, so installing the plugin brings the deja binary if you do not already have deja on PATH (via brew install deja-vu or install.sh). Run deja index once to index existing transcripts on disk, or run deja install --auto to also wire native MCP tools (recall, fix, blame, how, orient, context, remember, handoff) and lifecycle hooks across all 48 supported coding agents on this machine:
curl -fsSL https://raw.githubusercontent.com/vshulcz/deja-vu/main/install.sh | sh
deja install --autoIf a client already injects <deja-memory> or <deja-recall> through native hooks wired by deja install, the gateway hook detects those markers and skips duplicate injection.
- Local index lookup (
deja hook-prompt --plain): OnonRequest, the hook queries the local BM25 index under~/.cache/deja/index.db(DEJA_INDEX_DIRoverrides) asynchronously viadeja hook-prompt --plain(default timeout800ms). When a relevant session is found, the recalled excerpt is wrapped in<deja-memory>and appended to the current user turn — which means recalled excerpts are sent with the outbound request to your configured upstream model provider. - Strict project scoping (
ctx.cwd): By default (allowUnscoped: false), the hook requires a non-empty client working directory (ctx.cwd) and never falls back toprocess.cwd()(the Magpie gateway process's own directory).deja hook-promptfilters the index strictly to sessions recorded in that project so unrelated repositories' history is not mixed in. - Origin gate (
recallRemote: false): Only local requests (ctx.origin === "local") and authenticated owner requests (ctx.origin === "magpie-key") trigger recall unlessrecallRemote: trueis explicitly enabled. - Two-layer secret scrubbing:
dejaredacts common secret and token formats (sk-*,ghp_*,github_pat_*,AKIA*,Bearer,PRIVATE KEY, URL credentials,api_key=...) at index time, anddeja.hooks.mjsruns a second regex pass (maskSecrets) before attaching<deja-memory>to the request. - Duplicate guard & sticky turn cache: Requests that already carry native
dejarecall markers (<deja-memory,<deja-recall>,[deja —) are skipped. Recalled blocks are cached per conversation turn in memory so multi-turn conversations and tool loops keep earlier turns byte-identical, preserving the upstream provider's prompt prefix cache.
Configure in Magpie's plugin options for @vshulcz/magpie-deja:
| Option | Default | Description |
|---|---|---|
requestTimeoutMs |
800 |
Subprocess timeout in milliseconds (50–5000). |
maxBytes |
4000 |
Maximum UTF-8 byte budget for each <deja-memory> block (256–32768). |
recallRemote |
false |
Also recall on remote / gateway-key requests. |
allowUnscoped |
false |
Allow global recall when the client does not send ctx.cwd. |
cwd |
"" |
Explicit fallback project directory when ctx.cwd is absent. |
skipAgents |
[] |
Agent names (ctx.agent) to skip (for example ["claude-code"] if native hooks are already active). |
bin |
"" |
Explicit path to the deja binary. |
| Recall | Tokens | Index & lookup |
|---|---|---|
| 97.2% R@5 on LongMemEval-S (88.9% hit@1, 470 questions; MemPalace publishes 96.6%, agentmemory 95.2%) | 58% fewer tokens (half the tokens of agentmemory on repeated repo tasks) | 13.6 s to index 19,195 sessions / 0.7 ms lookup across 48 coding agents (next of 7 tools: 72 s) |
Share vetted project decisions and dead ends through git in .deja/memory.jsonl (deja repo candidates, deja repo add <id>). Entries are redacted on write, stored sorted with merge=union so parallel branches merge cleanly, and loaded for any teammate who clones the repository ({"activations":{"auto":{"repo":false}}} in the trust policy disables hook injection). More.
Switching tools mid-task? Run deja handoff --to <agent> (or use handoff via MCP) to transfer the active session's goal, decisions, touched files, passed checks, and open work into another coding agent.
Part of deja-vu. MIT. Found it useful? Star deja-vu on GitHub.
