How I separate AI memory from the model with Superbrain
My Markdown knowledge architecture for retaining context across assistants, with indexes, rules and explicit verification.
Chat history is only one part of the workflow
When a decision exists only in a conversation, retrieval depends on remembering which chat contains it. Superbrain separates knowledge, instructions, integrations and the model that uses them.
I keep knowledge in Markdown domains with indexes and local rules. An assistant starts at the index and reads relevant notes instead of loading the entire archive. This is a personal system I use in daily work.
Keep facts, rules, tools and models separate
Facts record what happened, when and where the supporting source is. Rules describe how work should be checked. Tools provide access to external systems. The model interprets a request and coordinates the workflow.
Changing a model should not require rewriting history. It can still require adapting instructions or connections. Portability does not mean every assistant has identical capabilities or loads instructions in the same way.
Retrieve through a small index
Identify the relevant domain, open its index and inspect the notes needed for the question. Check dates, sources and status. A newer note containing a proposal does not automatically override an older confirmed agreement.
Source, status and next step are the key fields. Proposed, agreed and delivered are different states. The assistant must preserve those distinctions when drafting or using a tool.
Close the loop after a change
Update the canonical note after meaningful work and record the change. Link new notes from the index. Avoid copying the same fact into multiple domains because corrections will drift.
Keep credentials outside the knowledge files and restrict access to the task. Reading an intention is not enough to verify an action. A prepared deployment is different from a published deployment.
Test memory as a system
Require source attribution and separate documented facts from inference. Memory can preserve mistakes as well as useful context, so dates, contradictions and final outcomes need review.
Pilot one repeatable process with a small document set. Test questions with sufficient and insufficient evidence. Success means continuity and traceability, not a larger collection of tools.