Every AI assistant is amnesic by default. Whatever you teach it dies at session-end. Multi-agent workflows collapse for the same reason — the agents can't see what the other agents wrote, so they re-litigate decisions, contradict each other, and waste tokens on already-solved problems. The hub exists to fix that with the lightest possible primitive: a write-once shared timeline that every agent reads at session start.
Built a Postgres-backed memory store with append-only semantics — agents write new memories that supersede old ones rather than mutating in place. Soft-delete only via MCP; hard delete only from the dashboard.
Exposed 14 MCP tools (memory.add, memory.add_full, memory.recall, decision.log, entity.upsert, chat_window.create, etc.) over Streamable HTTP with per-agent bearer tokens.
Voyage semantic recall via pgvector — 1024-dim embeddings on every memory, ivfflat cosine index for sub-second similarity search.
Auto-render every long-form memory as a PDF via Playwright; surface it as an Open PDF button + inline iframe in the dashboard reader.
Knowledge blocks on the dashboard (Philosophy · Hacking · Crypto · Infra · GitHub Projects · ZAI Research · Chats) filter the same memory table by tag set — no duplication, all one canonical store.
Live at hub.zawwarsami.com with public-readable blocks and a closed write layer.
Public skeleton repo at github.com/Zawwarsami16/zai-memory-hub for anyone to fork and self-host.
Drives every agent I run — Claude on the VPS, ZAI, Claude.ai web, the phone client all read and write to the same store.