TL;DR
- The language model itself does not carry a personal history between calls; the surrounding application supplies context.
- Products may call conversation history, saved facts, project files or searchable long-term records “memory.”
- Good memory is not maximum recall. It is relevant recall with clear boundaries and a way to correct or delete bad context.
- Test a product over time with decisions and rationales that you can verify.
Memory is an application capability, not a feeling created by a fluent answer. A model processes the context supplied for a call. The product around it decides what prior messages, saved facts, files or retrieved records to include.
That distinction explains why two products using similar models can remember very differently.
Four systems commonly called memory
Conversation context keeps the current exchange coherent by including recent messages. It is limited by the application’s context policy and should not be assumed to persist across a new chat.
Saved preferences or facts preserve selected items such as a writing preference or location. They are useful for personalization but do not form a complete work history.
Project knowledge retrieves from files, notes or a bounded workspace. It can provide durable reference material without remembering every interaction.
Searchable interaction history stores and retrieves prior conversations, decisions and events. This is the most relevant category for ongoing delegated work, but it can also retrieve stale or incorrect material if the system lacks good boundaries.
Products can combine these approaches. The useful question is not “does it have memory?” but “what is stored, how is it selected, and what control do I have?”
A practical memory test
- Record a decision, the alternatives considered and the reason for the choice.
- Start a separate session after enough time has passed to avoid relying on the current chat.
- Ask for the decision first, then its rationale and source context.
- Correct one detail and confirm the correction is used later.
- Introduce a newer decision that supersedes the first and test which one wins.
- Ask how to delete the test material and verify the documented behavior.
Score accuracy, relevance, stale-context errors and invented details. A system that retrieves more text is not necessarily better than one that retrieves the right evidence.
Brainmox’s current boundary
Brainmox stores conversations, memory, knowledge and uploaded files in a data store isolated for each customer. Its agent can search persistent records and use them across supported channels. That data is cloud-hosted; the optional local runtime applies to hands-on tools, not the memory system or model inference.
Memory does not guarantee proactive work, perfect recall or an outcome. Scheduling and Mission Control are separate capabilities that must be configured. Users should still verify recalled facts before using them in consequential work.