The prologue already held the hierarchy
Context Cascade’s four levels, queried one after another, are a cache hierarchy. They get the vocabulary a year later with the APU, and the metadata search becomes BMJ.
LLMs don’t have a memory problem. They have an architecture problem.
An open research series in a prologue and eight parts. Each part publishes the architecture before the code. Then the code and the measurements correct it. Every stage borrows from an architecture that has already proven itself, without copying it whole.
Memory track, 8 stages
What an agent is allowed to know
From a cascade of memories to a hierarchy with paging, offline tiers, and memory the person can read and delete.
- 2025-04-01Context Cascade Architecture
- 2026-01-20User-Centric Context Engineering
- 2026-03-15UX-Driven Agent Memory
- 2026-04-12Agent OS
- 2026-04-17Agent Processor Unit
- 2026-08-16Akili
- 2026-10Urithi
Context track, 6 stages
How agents are born, promoted and retired
From versioned blocks that developers assemble to agents that grow from usage and are retired on evidence.
- 2025-08-11Context as a Service
- 2026-01 → 04Borrowed from the memory track
- forthcomingContext Switch Engine
- 2026-07-16Context Lifecycle Engine
- 2026-07-17Build, Ship, Run
- 2026-09-06The cost of trusting AUC
How the systems work
Follow the flow, step by step
Urithi, in progress
A rule from the guardian’s voice, a question from the learner
Evidence, working memory, the debrief, the teach-back, the library, the anonymous question box, the signed offline package.
Two flows, animatedContext Lifecycle Engine
A Monday habit becomes an agent
Detect, build, run, promote on evidence, and demote when the model drifts.
Twenty steps, animatedThe series
Every article, in order
- Prologue
Context Cascade Architecture
memoryThree levels of memory in cascade, and blocks searched on their metadata. Three articles the same day.
- Part 1
Context as a Service
contextAgents as assemblies of versioned, tested blocks, promoted from dev to prod.
- Part 2
User-Centric Context Engineering
memoryOpaque, user-controlled, hybrid: context as a product feature.
- Part 3
UX-Driven Agent Memory
memoryThe Living DLL and the Bidirectional Metadata Jump.
- Part 4
Agent OS: what von Neumann knew
memoryLayers, not bigger registers, and a human interrupt.
- Part 5
From Agent OS to Agent Processor Unit
memoryA four-tier hierarchy, LRU paging and a priority scheduler. 78% L1 hit rate.
- Part 6
Context Switch Engine
contextA Merkle store of context states, and a switch that loads only the diff. In preparation.
- Part 7
Context Lifecycle Engine
contextThe system detects, creates and retires its own agents from how you use it.
- Part 8
Build, Ship, and Run Your Synthesised Agents
contextBuild, images, containers, volumes, and topology.yaml.
- Implementation
Implementing the APU: a tutor that runs offline
memoryAkili on salvaged hardware. The local tiers are 0.9% of the turn.
- Measurement
BigQuery, Gemini embeddings, and the cost of trusting AUC
contextAt a 1% error budget the free baseline goes from last to first.
Read in order
What the series rediscovers
It does not only move forward. It comes back to its own ideas, sometimes loses them, and finds them again by measuring.
The user never waits for memory
Set by Agent OS and the APU scheduler. Lost in the Akili rewrite, where extraction takes a third of the turn. Found again by measurement, then priced at 34.6% of perceived latency. Still an open decision.
Urithi returns to UCCE
Memory scoped to a context, a source for every fact, memory the person can see: all of it was in User-Centric Context Engineering. TencentDB Agent Memory was a reminder, not a source.
Method
Proved, not claimed
Theory first
The architecture is published before the code, to be defended before anyone is attached to it.
Code the next day
Article on 15 March, repository on 16 March. Article on 17 July, repository on 18 July.
Every number measured
On the real implementation, with its date, commit and model, and the command that reproduces it.
Say what it doesn’t show
Each measurement article ends on its limits. The tests carry hundreds of deliberate defects they must catch.