#Agentic Cognitive Architecture - Fight the model! Keep your agents from drifting

1 messages · Page 1 of 1 (latest)

narrow spoke
#

Agentic Cognitive Architecture (ACA)

A layered file-based cognitive architecture that fights coherence window decay. Your agent drifts because nothing in the architecture fights it. ACA provides the structural reinforcement that's missing.

What it is:

Six files on an abstract-to-deterministic spectrum (Soul, Job, Motivations, Charter, Comms, Identity)
Key mechanism: drift corridors — named behavioral failure modes that make the model's cheap default completions structurally expensive
Bootstrap hook that injects the extra files into agent context alongside standard OpenClaw workspace files
Works with any model
What's in the repo:

Full architecture README with research context, real production examples, and honest discussion of what it does and doesn't do
SKILL.md that your agent can read to help you get started
cognitive-stack bootstrap hook (handler.js + docs)
Apache-2.0. Start with SOUL.md. Add layers when you feel the gap.

https://github.com/PsiClawOps/AgenticCognitiveArchitecture

gilded quest
#

sounds interesting! But is it correct to hook the files into the bootstrap hook? Isn't bootstrap only applied once at first start of the agent?

narrow spoke
#

On startup. We added a hook plugin to expand the whitelist

#

My identity files have a health check procedure, does a full path post of the system files and states the model

narrow spoke
#

@sleek elk more here including how to do it

#

I actually had a bug in my first council session after inventing a cursor to have them take turns, I forgot to put an “end loop” in so they kept going lol

#

And they were like “wait he closed the floor why are we still talking” and “yeah we should be done by now” and some of them started troubleshooting the code as the turns kept going

#

“We must still be going because of a bug, a stop must have never been defined”

manic lily
#

I would love to see this in action to be honest

#

I am thinking to apply it to a use case i have, but any chance I get to see this applied by someone else on a use case will be totally awesome

narrow spoke
#

im going through a ton of testing with a research bot i made, trying to find out whats what, but i can tell you it definitely holds its identity

#

i use a lot of sonnet and opus and i have some agents ive talked to for thousands of turns and they still act the same