JSPM

Your agents forget everything. Phren doesn't.

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.

Install

npx @phren/cli init

One 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.

What actually happens

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.

Key features

Fragment graph

Explore connections visually. Drag nodes to reorganize; graph auto-settles. Click a fragment to see every finding linked to it across all projects.

Finding lifecycle

• 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.

Multi-agent support

Same store works with Claude Code, Copilot, Cursor, and Codex. Agents tag findings with their tool, so you see who discovered what.

Review queue

Mark findings as needing review ([Review] section). Phren surfaces review items on every session start. Approve, reject, or edit in place.

Governance & policies

Per-project retention policies. Confidence decay curves. Access control. Audit logs. Configure with phren config or the web UI.

Store subscriptions

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-projects

Progressive disclosure

Enable 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.

Semantic dedup & conflict detection

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.

Skills & hooks

Drop custom slash commands into ~/.phren/global/skills/. Hooks run on user prompt, tool use, and session events — wire phren into your own workflows.

CLI quick reference

phren                                   Interactive shell (explore/search)
phren search <query>                    Full-text search with FTS5
phren add-finding <project> "insight"   Capture a finding
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>   Create a team store
phren team join <url>                   Join a team store
phren web-ui [--port 3499]              Launch the web UI
phren doctor                            Health check & auto-fix

See full CLI docs at alaarab.github.io/phren.

Team stores

Shared knowledge repos for teams. One person creates with phren team init, others join with phren team join. Findings, 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.

Platforms

• 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.

MIT License. Made by Ala Arab.