vshulcz/deja-vu: claude Code, Codex, opencode, aider, Gemini CLI, Cursor, Antigravity, Grok Build

Your agents already solved this. deja finds it.
Search 3.3 GB of agent history in ~12 ms — and reach working context for ~60× fewer tokens than replaying it. One zero-dependency binary, fully local.

vshulcz.github.io/deja-vu

Claude Code, Codex, opencode, aider, Gemini CLI, Cursor, Antigravity, Grok Build, Qwen Code and pi write every conversation to local files — gigabytes of debugged problems and design decisions you can't search. deja is a zero-dependency binary that turns those histories into a memory layer:

Privacy

deja forget removes matching sessions from a rebuilt index and records exact session tombstones so a later deja index cannot restore them from source history. Tombstones are stored at ~/.config/deja/tombstones (or $XDG_CONFIG_HOME/deja/tombstones); use --dry-run, --list, or --unforget. Ingest exclusions are one case-insensitive project pattern per line in ~/.config/deja/exclude (XDG-aware), or comma-separated in DEJA_EXCLUDE_PROJECTS. deja stats --redaction reports redactions by harness and rule, along with tombstone and semantic-sidecar facts.

One binary. No models to download, no services to run, nothing leaves your machine unless you sync or share it. (opencode and Cursor IDE indexing shell out to the sqlite3 CLI, preinstalled on macOS and most Linux distros; Cursor CLI transcripts do not need it.)

Install

curl -fsSL https://raw.githubusercontent.com/vshulcz/deja-vu/main/install.sh | sh

or:

go install github.com/vshulcz/deja-vu/cmd/deja@latest   # Go
npx @vshulcz/deja-vu "query"                            # npm, no install
brew install vshulcz/tap/deja-vu                        # Homebrew

Shell completion

deja completion bash >> ~/.bashrc
deja completion zsh >> ~/.zshrc
deja completion fish > ~/.config/fish/completions/deja.fish

Wire it into the agents you use (edits config, keeps a .bak):

deja install --all     # MCP recall for every agent it finds on this machine
deja install --auto    # same, plus session-start auto-recall where supported

On Windows, register the MCP server through the shell wrapper most stdio servers need there: cmd /c deja mcp (deja install writes this form automatically; use it if you wire configs by hand).

Install also writes user-level guidance for the harnesses it detects: Claude Code, Codex, Gemini CLI, Qwen, Copilot, and OpenCode use their corresponding guidance files (or the configured XDG_CONFIG_HOME). Re-run rewrites deja's skill or marked block without changing surrounding user content. Use deja install --all --no-guidance to opt out; Cursor and Grok have no documented user-level guidance location and are skipped.

Install reports whether it found local history and builds the first index immediately when history is present.

That's it. Next session, ask your agent:

have we dealt with jwt refresh rotation before? check your memory

— or with --auto, don't ask: the agent starts each session already knowing what you solved in that project.

CLI

$ deja "jwt refresh token"
[claude] api        · Jul 8 · 8f31c0a9 — 2 matches
  login started failing after refresh token rotation; jwt kid mismatch in tests
  fixed by reloading jwks cache after rotateKey and adding a clock-skew test
[codex]  web        · Jul 1 · b77d91e2 — 1 match
  refresh token cookie needed SameSite=Lax in local callback flow

Stemmed JSON searches set "stemmed": true and include the catalog variants used.

Search hits carry exact, close, or semantic confidence tiers; close hits include the matched variant and semantic hits include cosine.

Share your stats

Run deja stats --card to write a self-contained deja-stats.svg for a README or profile. The command prints an embed snippet; commit the SVG to your profile or repository if you want it there.

Run deja stats --html to write a self-contained, browsable deja-stats.html timeline. The HTML export embeds metadata only: dates, harnesses, projects, message counts, and already-redacted first-user titles; it never includes message text.

deja update is for standalone installs. Homebrew and npm installs update through the package manager.

Doctor JSON

deja doctor --json reports an explicit state for every check and exits 0 even when it finds a problem. Store states are ok, missing, empty, unreadable, or parsed-zero; the last state means session files exist but the newest file produced no sessions.

{
  "stores": [{"name": "claude", "state": "ok", "paths": ["/home/me/.claude/projects"], "files": 42}],
  "index": {"state": "stale", "path": "/home/me/.cache/deja/index.db"},
  "mcp": [{"name": "claude-code", "state": "wired", "path": "/home/me/.claude.json"}],
  "sqlite3": {"state": "ok"},
  "version": {"state": "ok", "current": "1.2.3", "latest": "1.2.3"}
}

Index states are ok, missing, or stale; MCP states are wired, not-wired, or config-missing. The sqlite3 state is ok or missing. Version state is ok, update-available, ahead, dev, or unknown.

Context piping without MCP:

claude "Prior context: $(deja ctx 'database migration')"

Before changing a file, inspect its history:

deja blame cmd/deja/main.go

Semantic recall (optional)

Semantic search is an opt-in layer for a local Ollama, LM Studio, or OpenAI-compatible embedding endpoint. Set DEJA_EMBED_URL and optionally DEJA_EMBED_MODEL, then run deja embed. Ollama defaults to nomic-embed-text; without a configured and reachable runtime, ordinary lexical search and MCP recall continue unchanged. --no-embed or DEJA_EMBED=off disables reranking for one invocation. Without an embedding endpoint, the semantic zero-result fallback does not exist.

The vector sidecar is stored beside the index as .vectors.bin, not in index.db. Float32 vectors cost roughly 4 KB per 1k messages for a 1,024 dimension model, plus a small record key. Embedding is local and can consume CPU, memory, and model-server time; it never sends raw source files, only the redacted indexed text truncated to about 2k characters.

Sync between machines

Point both machines at one shared folder (Syncthing, iCloud, a git repo — anything that moves files):

deja sync export ~/Sync/deja   # machine A: appends new batches since last export
deja sync import ~/Sync/deja   # machine B: picks up what it hasn't seen

Or skip the shared folder when the other machine is a ssh hop away:

deja sync ssh mini          # push new records to mini and import them there
deja sync ssh mini --pull   # fetch mini's new records into this machine

ssh mode uses your system ssh/scp and the deja binary on the remote (looked up on PATH, falling back to ~/.local/bin/deja).

Batches are plain JSONL, redacted on the way out. Import is idempotent, so keep the folder as an append-only log and run both commands from cron if you like. Records never echo back to their origin. --full re-exports everything regardless of watermarks — useful when adding a machine after old batches are gone. Synced sessions show up under imported:<project> in search, recall, and session-start auto-recall.

Teach your agent to remember

deja install --all wires up MCP recall (Claude Code, Codex, opencode, Cursor, Gemini CLI, Antigravity, Grok Build, Qwen Code, pi — aider has no MCP client, pipe deja ctx instead); deja install --auto does the same and adds session-start auto-recall where the harness supports it (Claude Code hook, Codex hooks.json, an opencode plugin — Cursor, Gemini CLI, Antigravity, Grok Build, Qwen Code and pi have no hook that can inject context, so MCP is their full install). To make the agent reach for memory on its own, add this to your CLAUDE.md / AGENTS.md:

Before debugging or re-implementing something, run `deja "<query>"` (or the
 MCP recall tool) — past agent sessions across Claude Code, Codex, opencode, aider, Gemini CLI, Cursor, Antigravity, Grok Build, Qwen Code and pi
are indexed locally. Cite what you reuse.

MCP tools

With --auto, a SessionStart hook also feeds the current project's recent memory in automatically — read-only, capped at 2KB, and it never delays or breaks agent startup. Because SessionStart also fires after every context compaction, the same memory is re-injected right after Claude Code compacts — and a PreCompact hook captures the transcript into the index beforehand.

Security

Subagent transcripts are skipped by default (they mostly duplicate the parent session); set DEJA_INCLUDE_SUBAGENTS=1 to index them. Files caught mid-write are handled safely — the torn tail line is picked up on the next pass.

Credentials are redacted at index time: AWS keys, generic api_key=/token= assignments, bearer tokens and raw JWTs, PEM private key blocks, provider tokens (ghp_, sk-, npm_, xox., AIza), and scheme://user:pass@host URLs. The value is replaced with [redacted:<kind>]; surrounding text stays searchable. deja sources shows per-store counts. Opt out with DEJA_NO_REDACT=1 (unsafe). deja share and deja sync export re-apply redaction on the way out.

The security model documents data flows, redaction limits, trust assumptions, and release verification.

Supported harnesses

Custom locations via DEJA_CLAUDE_ROOT, DEJA_CODEX_ROOT, DEJA_OPENCODE_DB, DEJA_AIDER_ROOTS, DEJA_GEMINI_ROOT, DEJA_CURSOR_ROOT, DEJA_CURSOR_CLI_ROOT, DEJA_ANTIGRAVITY_ROOT, DEJA_GROK_ROOT, DEJA_QWEN_ROOT, DEJA_INDEX_DIR. Each agent's own relocation variable is honored too: CLAUDE_CONFIG_DIR, CODEX_HOME, GEMINI_CLI_HOME, CURSOR_CONFIG_DIR, GROK_HOME, AIDER_CHAT_HISTORY_FILE, and XDG_DATA_HOME for opencode on Linux.

DEJA_RECALL=safe is the default: SessionStart recall stays in the current project, filters weak or duplicate results, prefers the last 90 days, and injects at most 2KB. DEJA_RECALL=aggressive searches across projects and raises the injection cap to 4KB. DEJA_RECALL=off disables SessionStart recall output.

Session format registry

The session format registry documents the observed store paths, record schemas, role mapping, timestamps, and compatibility notes for each supported harness. Synthetic fixtures keep those descriptions checked against the parsers.

Performance

Measured on a real corpus — 1,250+ sessions, ~3.3GB across three harnesses:

The index is incremental: when a session file grows, only that file is re-read.

Benchmarks

Run the reproducible recall benchmark with:

deja bench recall
deja bench recall --json

The synthetic set is currently saturated by lexical search (recall@5 1.00 at ~0.7 ms median), so it serves as a regression floor for ranking changes rather than a bragging number; CI fails if recall drops. The corpus generator and the relevance labels are ordinary reviewed Go — audit what "relevant" means before trusting any figure, ours included. With a local embedding endpoint up, the same command reports the hybrid column.

The context experiment is reproducible with deja bench context --json. It builds 30 seeded multi-session task chains and five negative controls, then compares deja-recall (the real index and SessionStart digest), full-history, naive substring grep, and cold context. Tokens are approximated as bytes/4. Coverage is the fraction of generator-defined ground-truth fact strings present verbatim; audit that generator before trusting the figures.

Run with the default seed (1):

Prior sessions in the generated corpus carry realistic log-noise bulk; without it the full-history arm looks artificially cheap and the comparison is meaningless. On this corpus the recall digest reaches the same fact coverage as replaying the full history for about 60x fewer tokens, and injects nothing on the negative-control chains where no prior fact is relevant.

How it works

Local inverted index in ~/.cache/deja: parse JSONL/SQLite stores → redact credentials → records.bin + token buckets → manifest.json tracks per-file state so repeat runs only ingest what changed. The MCP server, stats, share and sync all read the same index. Details: docs/ARCHITECTURE.md.

Privacy: indexing and search are local. Network is used only by deja update, deja sync ssh, and the deja doctor version check. Local files in, local cache out.

FAQ

Does anything leave my machine? Indexing and search are local. deja update downloads releases from GitHub, and user-invoked deja sync ssh transfers redacted batches through the system SSH client. Directory exports and shares go only to the destination you choose. See the security model for the full data flow.

How is this different from cass? cass is the kitchen-sink take on session search: 22 providers, Rust, optional semantic embeddings, a TUI. deja is the opposite bet — one small Go binary, pure lexical, eight harnesses, zero setup — plus the memory-layer pieces around it: auto-recall, redaction, share, sync.

And from MemPalace / Mem0 / Letta? Those are memory platforms: embeddings, vector stores, capture hooks or APIs that record going forward. deja has no capture step at all — it indexes what your agents already wrote to disk, including months of history from before you installed it. They can coexist.

What about secrets already in my logs? They stay in the original harness files (that's your agent's data), but they don't enter deja's index, digests, shares or sync exports.

What about Windows? Builds exist, CI runs the suite on Windows; macOS/Linux are the battle-tested paths. Field reports welcome: #9.

Can I exclude a project? Not yet — planned as --exclude (#8). Today you can point DEJA_*_ROOT at a filtered copy.

How do I wipe everything?

deja uninstall --all
rm -rf ~/.cache/deja