Hippocampus: Giving Bobo a Working Memory

So last July after I left the business I started building a way to document and systematise everything I know.

And by that I mean everything; all the concepts, all the frameworks, all the thoughts, all the snippets of randomness that is the world of building and launching businesses, and how to do it. This manifested itself in a monolithic markdown files directory that I would plug Claude Code into. This file directory would include a CRM of the people I was dealing with at the time.

I would like to think I had a great thing going but then I would need to remember to run manual prompts to update docs and indexes. And then I would need to prompt Claude Code every time to create a system prompt of sorts before any kind of work I would do.

So what's the next most logical thing to do? Custom Next.js app to be my slave. And so I did. I created a web app that had all my CRM data, files and thoughts, mostly now migrated away from markdown files and living in Supabase. The app also had a chatbot that would initialise every time with its own system prompt, dynamically based on which company or topic I would talk about.

It was actually starting to work really well. I had given the bot a name, Bobo, and I also gave it its own personality and self-learning; so he would update himself and all the data every night via 2 cron jobs.

But he was also driving me crazy. Every new session I would start, Bobo would wake up confused about what is actually relevant to the conversation I want to have with it, now.

No memory of what we talked about yesterday, what deals were in the pipeline, what commitments were made. I'd spend the first 5 minutes of every session re-explaining everything or copy pasting stuff.

I kept trying different approaches and eventually landed on modelling the memory around a literal brain.

Three layers 1) daily notes as raw memory that would get generated from chats and updates to the files, 2) a rolling synthesis as working memory, and 3) permanent memory for things that stuck around long enough to matter. Like the early memory implementation from Claude.

And then when I did this, the conversations with Bobo started becoming more useful. Bobo started waking up knowing what was going on.

But... as fun as it was building all this, I'm not a software engineer. I had initially had everything in markdown format in a GitHub repo that I would use Claude Code from, but then managing the cron jobs manually was a mission so I started building the Next.js app — which is a full time job in itself!

I was both annoyed and relieved when OpenClaw dropped. It was like dammit I should have just stuck with the markdown based file system and figured out the cron and I would have had claw before OpenClaw!

But anyway, with OpenClaw, I started migrating everything over. All the files and contexts. But also the patterns and workflows I've been building.

Hippocampus is that pattern turned into a plugin. One install, your agents get a daily-synthesised briefing injected into every session so i.e. what's top of mind, what's open, what's owed. Not a log dump. A briefing.

Over the coming weeks I'll be deploying other parts of the brain. Would love to hear what your agent memory setup looks like today!

openclaw plugins install @sacheeperera/hippocampus