Persistent memory for AI agents. Findings, tasks, and patterns live in markdown files in a git repo you control. No database, no vendor lock-in. Works with Claude, Copilot, Cursor, and Codex.
npx @phren/cli initOne command. Sets up ~/.phren, wires up MCP for your tools, installs hooks. Next time you open a project, context starts flowing automatically. On a new machine? Re-run init and you're back in sync.
When you open a prompt:
- Hooks extract keywords from your question
- Phren searches findings across projects (FTS5 full-text with semantic fallback)
- Relevant snippets inject into your prompt before you hit send
- You ask; Claude already knows the gotchas
When you discover something:
phren add-finding <project> "finding text"captures it with optional tags ([decision],[pattern],[pitfall],[bug])- Trust scores decay over time; decisions never do; observations expire in 14 days
- Findings link to fragments (named concepts like "auth" or "build") that connect knowledge across projects
Sessions:
- Mark boundaries with
session_start/session_end - Next session sees your prior summary, active tasks, recent findings, and where you left off
- Checkpoints track edited files and failing tests so you can resume exactly where you stopped
Tasks:
- Add with priority/section. Pin across sessions. Link to GitHub issues.
- Track completions and cross-project rollups.
Explore connections visually. Drag nodes to reorganize; graph auto-settles. Click a fragment to see every finding linked to it across all projects.
- Supersede: "Finding X is obsoleted by finding Y"
- Retract: "We were wrong about this; here's why"
- Contradict: "We have two findings that conflict; this is why"
Helps you reason about contradictions instead of hiding them.
Same store works with Claude Code, Copilot, Cursor, and Codex. Agents tag findings with their tool, so you see who discovered what.
Mark findings as needing review ([Review] section). Phren surfaces review items on every session start. Approve, reject, or edit in place.
Per-project retention policies. Confidence decay curves. Access control. Audit logs. Configure with phren config or the web UI.
Subscribe to specific projects in a team store — others stay hidden from search and context injection:
phren store subscribe team-store arc intranet
phren store unsubscribe team-store legacy-projectsEnable PHREN_FEATURE_PROGRESSIVE_DISCLOSURE=1 to get compact memory indices instead of full snippets. Call get_memory_detail(id) to expand only what you need.
Optional: enable LLM-based duplicate detection and contradiction flagging on add_finding. Prevents near-duplicate entries and catches "always use X" vs "never use X" contradictions.
Drop custom slash commands into ~/.phren/global/skills/. Hooks run on user prompt, tool use, and session events — wire phren into your own workflows.
phren Interactive memory shell
phren search <query> Full-text search with FTS5
phren add-finding <project> "insight"
phren note add <project> "daily note" Add a lightweight daily note
phren task add <project> "item" Add a task
phren session_start <project> Start a session
phren store list List personal + team stores
phren team init <name> --remote <url>
phren team join <url> Join a team store
phren web-ui [--port 3499] Launch the web UI
phren doctor Health check & auto-fixSee full CLI docs at alaarab.github.io/phren.
Shared knowledge repos for teams. One person creates with phren team init, others join with phren team join. Findings, daily notes, tasks, and skills sync across team members.
Each team store can be configured with per-project subscriptions so people only see what they care about.
- Claude Code (VS Code, Web, Desktop) — MCP hooks + CLI
- Copilot (VS Code, GitHub.com) — MCP hooks
- Cursor (IDE) — MCP hooks + built-in skill system
- Codex (Claude Agent SDK) — MCP tools + hooks
All use the same phren store. No vendor lock-in.
| Package | Description |
|---|---|
@phren/cli |
CLI, MCP server, data layer (59 tools, FTS5, hooks) |
phren-vscode |
VS Code extension (sidebar, graph, onboarding) |
MIT License. Made by Ala Arab.