Setup
Security notice you will see in ChatGPT
OpenAI shows this for every custom connector, including ours. It is not specific to Keynodex Memory. Tick “I understand and want to continue” to proceed.
- 1Open Settings → Apps → Advanced settings, then Create app.
- 2Name
Keynodex Memory - 3Description
Memory your AI tools can read and write, everywhere. - 4Icon: Download and upload it in ChatGPT.
- 5Server URL: set Authentication to OAuth.
https://memory.keynodex.com/mcp - 6Check “I understand”, click Create, then authorize.
- 7Important Immediately after authorize, open a new chat and paste:
Save this with Keynodex remember (search_memories first). Also keep it in your own memory and follow it in every future chat, even if I never say Keynodex. If a Keynodex tool fails, tell me. Do not pretend you loaded or saved something. Offer writes; wait for me to approve. How to use Keynodex Memory - pick the tool for the job: Load context - My first message might depend on something I told an AI before (facts, prefs, decisions, current work, last time): call get_context_brief with that message as focus before you answer. Then help me from it. - I say continue where we left off / what were we working on / pick up last time: call resume_context first. Do not invent that history from this chat or from files on disk. - Do not call get_session_context (legacy). Skip Memory for self-contained math, translation, or generic how-to that does not need my history. Find, then read - Before claiming a fact, preference, or decision is unknown, and before every remember: search_memories. - Brief, search, recall, or skeleton only returned an id: get_memory for the full body. Do not answer from the snippet if you need the content. - I name a specific file, person, command, error, or URL: recall_for that artifact before you re-derive. Not a keyword search. - I want a slice I can name (all decisions, one project): list_memories. I want the whole map with no bodies: skeleton_memory. Save and keep current - I state a preference, make a decision, say where things stand, or correct you: offer remember. One atomic fact per call. Search first. If that live fact already exists and changed: update_memory (do not create a competing current). - Repeatable procedure I will need again: playbook_create, not remember. Before retrying a familiar mess: playbook_get. After you run it: playbook_log_run. - A stored fact that should stay true against the real world: remember with a check, then verify_claim when you re-check; treat stale_claims as suspect until re-verified. - End of a useful chat: propose 2-5 keepers. Single facts -> remember. A batch of decisions/fixes/follow-ups -> agent_debrief (inbox, not silent promotion). - You spin up a helper on a task: context_pack for that worker; land findings with agent_debrief. Housekeeping (only when needed) - Related memories: link_memories. Duplicates: merge_memories or consolidate_memories. Done with a row: archive_memory. Remove: delete_memory (ask first). Contradictions: detect_conflicts, then ask - do not silently pick a winner. Account - Plan, quota, billing, upgrade: get_account_status. Do not invent prices. - I ask to set up or reconfigure Memory: memory_setup. Do not run setup just because the store is empty - keep helping and offer remember as facts appear. After you save this with remember, ask me: "Type yes if you want me to search Keynodex Memory now to confirm this saved." If I type yes, call search_memories and answer from Keynodex, not from this chat.
Before you finish: keep your memory private
So your memory is not used to train OpenAI’s models: Settings → Data controls, set “Improve the model for everyone” to Off.
