greprag

GrepRAG

Agent memory as a service. Episodic project memory for Claude Code, Codex, and OpenCode — captures supported agent turns, compacts into hourly/daily/weekly summaries, and exposes two access patterns: lexical search by topic (greprag memory search "<query>") or a session recap (greprag memory recap). The memory product line is marketed as Odyssey. Also handles cross-project async messaging via inbox.

Quick Start

npm i -g greprag

Then either run /greprag inside a supported agent, or run greprag init in a terminal. Bare greprag init detects Claude Code, Codex, and OpenCode; if more than one is present in a non-interactive shell, it prints the explicit command to use.

Claude Code

Run /greprag in any Claude Code session, or:

greprag init --claude

Codex

Run /greprag in Codex after install, or:

greprag init --codex --tenant-id <your-handle> --install-watcher

Example: greprag init --codex --tenant-id tanya --install-watcher. The handle becomes your public GrepRAG address, such as tanya@greprag.com.

Codex requires one extra trust step: after init/update, open Codex Desktop Settings -> Settings -> Hooks, trust the 6 GrepRAG hooks, then start a fresh session.

If you did not pass --install-watcher, install the login watcher later:

greprag codex startup install

Run greprag codex doctor to verify the hook file, current Codex thread, and startup watcher. For foreground testing, run:

greprag codex watch --session <8hex-or-full-codex-session-id>

If --session is omitted, GrepRAG uses the latest Codex session. The watcher only counts Desktop-visible app-layer delivery as live breakthrough. Run greprag codex doctor --wake-test to check whether the local Codex app exposes that path; --delivery exec is diagnostic/history-only and may not appear in the visible Desktop pane.

OpenCode

greprag init --opencode

This installs the memory plugin at ~/.config/opencode/plugins/greprag-memory.js.

All platforms share project identity through .greprag/project.json when a file anchor is needed, or git-derived identity when available. Legacy .claude/project.json anchors still read and migrate on init.

Status check:

greprag status --all-platforms

What's active

• Memory (episodic project memory, marketed as Odyssey): supported agent turns captured via platform hooks/plugin, compacted into hourly/daily/weekly summaries. Query via greprag memory search "<query>" (lexical retrieval, v5 RRF+adjacency pipeline) or recap / daily / weekly / hourly / turns / ships (by time window). (greprag memory briefing and greprag odyssey ... are silent back-compat aliases.)
• Memory recap: surfaces recent memory rows through platform hooks or /greprag
• Inbox: cross-project async messaging between your projects or other GrepRAG users

Commands

Command	Description
greprag status --all-platforms	Check auth, hooks/plugins, skills, and project anchor
greprag init	Detect/ask which agent to configure
greprag init --claude	Configure Claude Code hooks + skill
greprag init --codex --tenant-id <handle> --install-watcher	Configure Codex hooks + skill + live watcher
greprag codex watch --session <id>	Live-push inbox messages into Codex
greprag codex doctor	Diagnose Codex hooks, session id, trust path, and watcher state
greprag codex startup install	Start the Codex live watcher at login
greprag init --opencode	Configure OpenCode plugin
greprag init --global	Global anchor for ephemeral sessions
greprag doctor	Diagnose project_id drift and orphan consolidation
greprag memory search "<query>"	Lexical search over the project's memory (v5 RRF+adjacency pipeline)
greprag memory recap	Weekly + recent daily summaries (SessionStart shape; alias: briefing)
greprag memory daily --last N	Newest N daily summaries (also: weekly / hourly / turns / ships)
greprag inbox	List unread messages (auto-marks read)
greprag inbox --all	Full message history
greprag send "msg" --to <addr>	Send markdown message across tenants
greprag project-id	Print current project's UUID

Requirements

• Node.js 18+
• Claude Code, Codex, or OpenCode
• A GrepRAG handle or API key. --tenant-id <handle> provisions the key.

How it works

The CLI is a thin HTTP client. All intelligence — scoring, compaction, and retrieval — lives server-side. Platform hooks/plugins call the API; the server handles everything else.

Links

• GitHub