Akili
A school tutor for classrooms where the connection comes and goes. Only the model call leaves the machine, and the memory tiers take 0.9% of a turn.
I design how AI agents remember, and how they evolve.
Ezekias Bokove, Applied AI Engineer and Google Developer Expert for Google Cloud.
The Agent Processor Unit is my open architecture for agent memory. It runs memory the way a processor does: a cache in RAM, a page table, paging to disk, and a scheduler that answers first and writes later. Every number behind it is measured on a real system.
Built on the Agent Processor Unit
Two systems run on it, and one engine grows agents around it.
Urithi
An elder teaches by talking and showing. Urithi structures it, tests it like code, and hands it on to each learner in the form they need, online or offline.
Context Lifecycle Engine
Watches how someone works, and turns a habit into an agent that is promoted on evidence and retired when its proof expires.
How the architecture got here: eight stages of memory, from a 2025 chatbot that forgot too much, and six stages of agent lifecycle, from versioned blocks to agents that grow from usage.
Writing
Recent articles
Context Lifecycle Engine: BigQuery, Gemini Embeddings, and the Cost of Trusting AUC
Google CloudAt a 1% error budget, free word overlap goes from last to first. Benchmark your operating point, not your model.
Implementing the Agent Processor Unit: A Tutor That Runs Offline
MediumAkili on salvaged hardware: only inference leaves the machine, and the local tiers are 0.9% of the turn.
Context Lifecycle Engine: Build, Ship, and Run Your Synthesised Agents Anywhere
Google CloudThe YAML is a Dockerfile, the approved agent is an image, the running agent is a container.
Context Lifecycle Engine: When the Agentic System Detects, Creates, and Retires Its Own Agents
Google CloudReformulation triggers a repair, recurrence an automation, and every promotion needs evidence.
Passwordless Between Google Cloud and MongoDB Atlas
Google CloudWorkload Identity Federation from trust anchor to live write, with a reproducible demo.
From Agent OS to Agent Processor Unit
Google CloudA four-tier memory hierarchy with LRU paging and a scheduler that keeps replies instant.
Implement secure CI/CD with WIF, GitLab CI and Cloud Deploy
MediumGitLab CI for integration, Cloud Deploy for delivery, no long-lived keys.
Database migration using Cloud Seed
MediumA CI pipeline that migrates Cloud SQL schemas with GitLab’s Cloud Seed.
Speaking
40+ talks, workshops and webinars
DevFests, Google I/O Extended and community events across Africa and online. Open an event to see its photos and the talk.
About
Ezekias Bokove
I’m an Applied AI Engineer and Cloud Database Engineer, and a Google Developer Expert for Google Cloud. I founded GDG Cloud Abidjan and the GDGs Afrique Francophone network.
I care about AI, cloud, databases, APIs and DevOps, and about projects that improve people’s living conditions. That is why my agents are built for the floor first: a system that fits a salvaged laptop in a classroom also fits a server, and the reverse isn’t true.
- talks and events
- 40+
- published articles
- 30+
- parts in the Agent OS series
- 8