Tages
Team memory for AI coding agents.
Your AI agents forget everything between sessions. Every decision re-litigated. Every convention re-explained. Every past mistake repeated by the next agent that touches the same code.
Personal memory tools patch the symptom: store some facts, retrieve them later. That works for solo developers who remember what they stored. It breaks down when teams ship with AI — when three developers are using agents on the same codebase, when onboarding requires transferring institutional knowledge, when the wrong memory survives a refactor because no one audited it.
Memory isn't storage. It's a team practice.
Tages treats codebase memory as a managed artifact: structured types, quality scoring, audit trails, sharpen passes that rewrite vague notes into imperative agent instructions. One developer's architecture decision becomes every agent's context. A bad memory gets flagged and corrected before it misleads the next session.
Start in under 60 seconds with one command. Add team features when your workflow demands them.
claude mcp add tages -- npx -y @tages/server
Why Tages?
| Tages | Zep | Mem0 |
|---|
| Install | One line (claude mcp add) | Docker + API key | API key + SDK |
| Local-only mode | Yes (SQLite, zero config) | Self-hosted only | No (cloud-only) |
| Team sharing | Yes (RBAC, federation) | Yes (cloud) | No |
| Dashboard | Yes (Next.js, analytics) | Yes | Basic |
| Quality control | Audit, sharpen, enforce | No | No |
| Memory types | 11 structured types | Knowledge graph (Graphiti) | Structured |
| MCP tools | 56 | N/A | N/A |
| Search | Trigram + semantic + decay | Temporal knowledge graph | Vector |
| Workflow integration | Git hooks, CI/CD, briefs | SDK calls | SDK calls |
| Pricing | Free local / $14 Pro | Open-source / Cloud | $19-$249/mo |
What It Remembers
Source code tells agents what exists. Tages tells them why it was built, how to work with it, and what not to do.
| Type | Example |
|---|
| Convention | "Always use snake_case for API routes" |
| Decision | "Chose Postgres over MongoDB for pg_trgm fuzzy search" |
| Architecture | "Auth middleware in lib/auth.ts, JWT in httpOnly cookies" |
| Lesson | "Don't cache the Supabase mock — tests need fresh state" |
| Anti-pattern | "Never pass id in upsert with onConflict — causes FK violation" |
| Pattern | "All API errors return { error, code, status }" |
Plus: entity, execution, operational, environment, preference (11 types total).
How It Works
- Install —
tages init connects to your project, or install the Claude Code plugin for zero-config setup
- Remember — Store decisions and conventions manually, via git hooks, or by importing CLAUDE.md
- Recall — Every session gets full project context in <10ms from local SQLite cache
Zero-Config Auto-Detection
When running as a Claude Code plugin or MCP server, Tages automatically detects which project you're in:
.tages/config.json — explicit marker file (created by tages link)- Git remote — matches the repo name against registered projects
- Directory name — matches the folder name against registered project slugs
- Auto-create — if authenticated, creates a new cloud project automatically; otherwise uses local-only mode
No tages init required per directory. Use tages link [slug] to explicitly bind a directory to a project.
Claude Code Plugin
Install Tages as a Claude Code plugin for automatic session memory:
/plugin https://github.com/ryantlee25-droid/tages
Works With
Claude Code, Cursor, Codex, Gemini — anything that speaks MCP.
Features
- 56 MCP tools — remember, recall, audit, sharpen, import, federation, analytics, and more
- 53 CLI commands — full control from the terminal
- Web dashboard — browse, search, and edit memories with dark-mode UI
- Auto-indexing — git hooks extract decisions from commits via Ollama or Claude Haiku
- Import — seed from existing CLAUDE.md, ARCHITECTURE.md, or JSON files
tages brief — generate a cached context document for system prompt injectiontages audit — score your memory coverage and get suggestions for improvementtages sharpen — rewrite memories into imperative form for better agent consumption- Local-first — SQLite cache for sub-10ms queries, works offline
- Hybrid search — pg_trgm trigram matching + pgvector semantic search
- Team sharing — multiple developers share one codebase memory (Pro)
Benchmarks
Reproducible LongMemEval and coding-memory benchmark results are published under eval/ with full methodology, judge configuration, and run notebooks. Results are reproducible against the published harness; raw numbers are in each eval's results/ directory.
Setup Guides
Architecture
packages/
server/ MCP server (56 tools, stdio transport, 818 tests: 805 passing, 13 skipped)
cli/ CLI (45 top-level commands, npm global install, 331 tests)
shared/ TypeScript types + Supabase client
apps/
dashboard/ Next.js 16, Supabase Auth, Tailwind, shadcn/ui
supabase/
migrations/ 64 migrations (tables, RLS, pgvector, RBAC, encryption)
This is the beginning of the README. Read the rest at the original. Read the full README ↗
