bug: bot forgets / disregards preferences in "Knowledge Context"

[ats-bcon]

Requirement: The user "preference" memories should be of higher priority for the LLM. Here is why:

I have made the bot create a memory "User wants every response to include a summary of operations performed as the last part." in Knowledge Context but it does not follow it. (see screen-shots). So what is the value of "Knowledge Context" if the bot does not use it anyway and tells me to write memories to ROLE.md !? The whole purpose of the memory system is for the bot to use it, isnt it? Maybe session instcutions should be changed for the LLM to prioritize and always follow "preference" memories.

Memory entry:

Chat: (I have made the bot reread the memories. Only after that it started to follow the preferences):

[jamiepine]

It telling you to add it to your ROLE.md is pointless, it's already in the Knowledge Context. Which model are you using? This is not a failure of the system as much as it is the model simply ignoring the context, some models are better than others at this, though I'm constantly improving the system prompts. Knowing your model will help!

[ats-bcon]

Sure. My model is claude-sonnet-4-6 with a direct / native anthropic api key. (This is prolly the second most expensive option on the market right now, opus-4-6 being the first)

[ats-bcon]

P.S. I have added the following prompt-snippet to spacebot\prompts\en\channel.md.j2 under "How You Work" to experiment if it helps.

"At the start of each new session, branch immediately to recall your preference memories — call memory_recall with mode: "typed" and memory_type: "preference". Read what comes back and let it shape how you engage for the rest of the conversation. Do this silently (skip after the branch returns). Do not wait for the user to prompt you."

[ats-bcon]

@jamiepine I probably found the error. The branch that recalls memories does not return the memories to the channel.

I added a prompt as described in the previous message and here is what I got:

Once asked, it found the memories:

[jamiepine]

I'm very certain branches return memories to the channel, thats their only job and works well in mine and others' experience. Though I haven't tested with your prompt. I wouldn't suggest branching at the start of the session either, the whole point of the cortex is to provide the context for the session ahead of time. If you want to tweak anything, it would be how the cortex synthesiszes the working memory and knowledge context.

[jamiepine]

I ate my words just tested, seems like a recent bug, branch didn't retrigger channel llm turn, working on a fix

[ats-bcon]

@jamiepine , one ore thing ...

before extending the prompt I just talked to the "channel" and observed what went out to the model. Here is what I saw. The memories were not injected.

P.S. Please dont eat your words :)

[ats-bcon]

here is another hint that branching has issues with memory_recall: