Lite fixed the problem for one developer. This is what we built when the team joined — PostgreSQL + pgvector + n8n so that decisions made in one session, by anyone on the team, are visible everywhere within seconds.
Lite solved the solo problem. When more people joined, the same failure mode multiplied.
Engineer A in London writes a requirement. Engineer B in Singapore starts building the opposite. No one knows until code review — days later.
The same architectural question gets answered five times across five sessions because no one can search what was already decided.
A developer leaves the team. Three months of decisions, error patterns, and schema knowledge walks out with them.
Hooks on every machine funnel into one central database. Every agent queries the same store.
Each developer installs Zikra's Stop and PreCompact hooks. At the end of every Claude Code or Cursor session, the hook fires and POSTs the session summary to your central n8n instance.
Your self-hosted n8n instance receives every session payload. It calls the OpenAI embedding API (or your local model), generates a 1536-dim vector for each memory, and writes both the text and the vector to PostgreSQL.
The most common failure mode in distributed teams: two people building in opposite directions. Zikra Pro closes that gap.
A product manager in one location writes a requirement:
"API must respond within 200ms under normal load."
It's saved with save_requirement and immediately indexed in the central PostgreSQL store.
Two hours later, an engineer in a different city starts a new Claude session. The session's context file automatically pulls the top requirements matching the current project — and Claude opens with full awareness of the constraint.
Zikra builds on the same API surface. No commands to relearn.
Issue separate bearer tokens for each developer, CI machine, or external agent. Role-based control: reader · writer · admin. Scope tokens to specific projects. Revoke instantly.
The active_runs table records every live session. See who is working on what, in real time. Session IDs extracted from Claude Code JSONL transcripts automatically.
IVFFlat index with 100 lists handles millions of memories without degrading search quality. PostgreSQL's battle-tested durability and WAL replication — your team's memory doesn't disappear.
n8n sits between hooks and PostgreSQL. Add custom logic: Slack notifications when a critical decision is saved, weekly memory digests, or auto-tagging by project keyword. All configurable without code.
Stop hook, PreCompact hook, session capture daemon, and the Neovim statusline plugin. Every touchpoint in a developer's workflow can feed Zikra Pro automatically.
Claude Code, Cursor, Aider, Gemini CLI — if it can fire a shell hook or call a webhook, it works with Zikra Pro. One memory store for every tool your team uses.
Three commands to bring up the full stack. One curl to verify.
Same API surface. Same command names. Same JSON shapes. Point your agents at the new endpoint and your existing integration keeps working without a single change.