One portable memory that every personal agent can read

Gonto built a personal wiki that lives outside any single agent so he can switch between Hermes, Grok Bot and Instinct without re-explaining himself. It is mostly Markdown, split into CRM, Notes, Articles and Family sections. A Nightly CRM job in Hermes reads the day's Gmail, Slack, WhatsApp, Granola transcripts and agent sessions, extracts asks, decisions and commitments, and deterministic code validates and applies the updates. Other agents reach it through authenticated MCP and REST with four operations: search, read, list and save. Why it matters: It's a concrete design for owning agent context instead of leaving it trapped in each product. The useful patterns are separating saved from searchable, hybrid SQLite search (BM25 plus embeddings fused by reciprocal rank), per-record embedding policies for sensitive data, and making agents read the source evidence after search instead of trusting the hit. Per-client revocable credentials, idempotent saves and keeping CRM writes inside one agent stop several agents from corrupting shared memory. One caveat is that this is one person's self-built setup, with no numbers on quality or cost.

Try this: Pick one context you re-explain to every agent, such as your key relationships or current projects, put it in a single Markdown store with search, read and save operations, and tell each agent to search it and read the evidence before asking you.