Your team uses different AI tools. None of them share memory by default. Zikra fixes that.
Zikra is a self-hosted memory layer that works across Claude Code, Gemini CLI, Codex, and any agent that can send a POST request. Decisions, errors, architecture notes, and requirements are stored once and surfaced automatically — to every agent, on every machine, for every team member.
One webhook. One token. One shared context pool. No cloud. No SaaS. Your data stays where you put it.
- Teams where one person uses Claude Code, another uses Gemini CLI, another uses Codex — and they're constantly re-explaining context to each other
- Solo developers who switch between AI tools and hate starting from scratch every session
- Any workflow where an AI agent needs to know what another agent decided last week
Claude Code ──┐ Gemini CLI ──┼──▶ Zikra (MCP + webhook) ──▶ PostgreSQL + pgvector Codex CLI ──┘ ↑ Any agent ────────────────┘ One memory pool. Every agent reads and writes it.
One POST to save. One POST to search. Hooks fire automatically at session end — you never type "save this." At the start of every session, every agent gets the same shared context injected automatically.
| Feature | What it does |
|---|---|
| Cross-platform memory | Claude Code, Gemini CLI, Codex, ChatGPT — same webhook, same token, same memory pool |
| Team shared context | Every developer on every machine reads and writes the same pool. No silos. |
| Role-based access | Owner, collaborator, guest — enforced at the router. Each person gets a scoped token. |
| Hybrid search | 70% semantic vector + 30% keyword. Finds by meaning, not exact match. |
| Confidence decay | Memories age automatically. Errors decay in 30d, decisions in 90d, schemas in 180d. No manual cleanup. |
| Requirements pipeline | Write a requirement in any AI. Owner promotes it. Developer pulls and executes it. Full traceability. |
| Auto-save hooks | Stop hook fires when any session ends. PreCompact hook saves before context resets. |
| Active runs table | See which machine is running which prompt in real time across your team. |
| Session tracking | Every run logged — tokens, cost, files modified, linked to the prompt that triggered it. |
| Schema storage | Store DB schema per project. Any agent pulls current version before touching backend code. |
| Self-documenting API | Unknown command returns full reference with examples. No external docs needed. |
| MCP native | Claude connects via MCP trigger. No custom integration. |
Zikra — PostgreSQL + pgvector + n8n (recommended for teams): Fetch https://raw.githubusercontent.com/getzikra/zikra/main/prompts/g_zikra.md and follow every instruction in it.
Zikra Lite — SQLite + FastAPI, ready in 60 seconds (great for solo or small teams): Fetch https://raw.githubusercontent.com/getzikra/zikra-lite/main/prompts/g_zikra_lite.md and follow every instruction in it.
Both variants have the same full feature set and identical API surface. The difference is infrastructure — SQLite vs PostgreSQL. Upgrade from Lite to Full is one URL change.
| Zikra | Zikra Lite | |
|---|---|---|
| Backend | PostgreSQL + pgvector | SQLite |
| Concurrent team writes | ✅ Unlimited | Limited |
| Scale | Billions of rows | Millions of rows |
| n8n workflows + scheduling | ✅ | ❌ |
| Automated decay via cron | ✅ | Manual |
| Setup time | 30–60 min | ✅ 60 seconds |
| All memory features | ✅ | ✅ |
| Full API surface | ✅ | ✅ |
| Free | ✅ MIT | ✅ MIT |
| Repo | Description |
|---|---|
| getzikra/zikra | Full stack: PostgreSQL + pgvector + n8n + MCP server |
| getzikra/zikra-lite | Lightweight: SQLite + FastAPI + MCP server |
| zikra.dev | Docs and onboarding |
One memory. Every agent. Every team member.