ChatGPT cannot reach a memory server on your laptop, and its own memory is not something you can search or delete record by record. This is the other shape: a remote endpoint you own, on infrastructure that costs nothing at small scale. The tool surface is large because the structure is — wings, rooms, drawers, diary, knowledge graph and tunnels — so start with the wake-context call and let its guidance lead.
A remote memory server you deploy yourself on Cloudflare Workers. Memories are kept as drawers and diary entries with a knowledge graph over them, searched by a hybrid of semantic and lexical retrieval, and scoped per tenant. Hosted clients connect over Streamable HTTP with OAuth.
- mempalace_wake_context — the call to start a chat with, returning bounded, scope-limited context
- mempalace_search and mempalace_check_duplicate — hybrid retrieval, and a duplicate check before you write
- mempalace_add_drawer, mempalace_update_drawer and mempalace_delete_drawer for durable notes, indexed as they are written
- mempalace_get_drawer, mempalace_list_drawers, mempalace_list_wings, mempalace_list_rooms and mempalace_get_taxonomy for browsing the structure
- mempalace_diary_write, mempalace_diary_read, mempalace_diary_search and mempalace_diary_reindex
- mempalace_kg_add, mempalace_kg_query, mempalace_kg_invalidate, mempalace_kg_timeline, mempalace_kg_check and mempalace_kg_stats — temporal facts, with a check for conflicts and staleness
- mempalace_traverse, mempalace_find_tunnels, mempalace_create_tunnel, mempalace_list_tunnels, mempalace_delete_tunnel and mempalace_follow_tunnels for the graph between locations
- mempalace_status, mempalace_reconnect, mempalace_hook_settings and mempalace_memories_filed_away for diagnostics and write status
- Tenant scoping, so separate people or workflows cannot see each other's memory
A Cloudflare account with Workers, D1, R2, Vectorize and Workers AI enabled, plus Node.js and an authenticated wrangler. You generate three secrets, mint an access key per tenant, and deploy — no VM, no Docker, no database to administer. There is no shared public instance: you run it. Free-tier limits apply and Vectorize in particular has its own constraints, so check the pricing before rolling it out widely. ChatGPT is the client that has been verified end to end.