What an agent is allowed to know

Prologue2025-04-01theory

Context Cascade Architecture

It started with a user’s frustration: repeating to a chatbot what it had already forgotten. The answer was to stop keeping everything in one place.

• A cascade: context window, temporary memory, persistent memory, external database. Queried in order, and only when the level above is not enough.
• Historical blocks with metadata, searched from the most recent, skipped on their metadata before their content is read, with a threshold to stop early.
• Redis for the temporary buffer, and any vector, SQL or NoSQL store for the archive.

Among the use cases: personalised tutors, and assistants for the elderly. Both come back at the end of this page.

The paradigm Temporary memory Persistent memory

Part 22026-01-20theory + code

User-Centric Context Engineering

The more autonomous an agent becomes, the less users trust it. Context stops being an implementation detail and becomes a product feature.

• The automatic cascade. An algorithm no longer decides alone what is kept.
• Inbox and Bank. The agent writes to an Inbox it can never read. Only facts the user approves reach the Bank, stamped approved_at.
• A glass-box dashboard, and agentic refactoring: change the memory with one sentence.
• Extraction moves to a fire-and-forget background worker.
• Three modes on the same stack: opaque, user-controlled, hybrid. The hybrid writes optimistically and lets the user correct afterwards.

A black box breeds mistrust, a gate breeds fatigue, a garden breeds collaboration.

Article EZFRICA/context-engineering

Part 32026-03-15theory + code

UX-Driven Agent Memory

The problem gets a name, context pollution, and memory gets a shape. The model is still human memory: working, episodic, semantic.

• A Living DLL decides what the agent may see: four fixed blocks always injected, up to eight dynamic ones the agent proposes and the user confirms.
• BMJ, the Bidirectional Metadata Jump: test the head, and below the threshold of its type, jump to the tail and walk back. The prologue’s most-recent-first search, made bidirectional.
• Letta stores, Weaviate indexes, the DLL decides.
• Inbox and Bank. Control moves from approving every fact to forcing or silencing whole blocks.
• Twelve blocks is the limit for a travel agent, not a universal setting.

Fixed blocks are a contract, Force Override is a right, BMJ is a service.

Article EZFRICA/ux-driven-agent-memory code the next day

Part 42026-04-12theory

Agent OS: what von Neumann knew

The analogy changes source, from human memory to the computer. In 1945 the answer was not bigger registers but separate layers.

• The Force (Override) checkbox becomes an interrupt: inspect the assembled context before the LLM sees it, remove or inject blocks, roll back a write.
• One memory shared by every interface: terminal, IDE, dashboard.
• Native multi-tenancy, one Weaviate tenant per agent.
• The write-back runs after the reply and is transactional: no layer knows something the others don’t.

Article same repository as part 3

Part 52026-04-17theory + code

Agent Processor Unit

The prologue’s cascade finally gets its real name, a memory hierarchy, and a processor’s units to run it.

• An L1 cache in RAM in front of the DLL. Weaviate becomes L3, Letta becomes L4.
• LRU paging at twelve active blocks, with page faults that bring an evicted block back.
• A scheduler with four priorities: the reply first, archive sync in the background with three retries, garbage collection when idle.
• A Tool Execution Unit that caches tool results for fifteen minutes.
• BMJ becomes the TLB: the DLL first, L3 only below 0.80.

Article EZFRICA/Agent-Processor-Unit

Implementation2026-08-16code + measurement

Akili, a tutor that runs offline

The prologue’s personalised tutor, sixteen months later, on salvaged hardware where the connection comes and goes. One rule: only inference may leave the machine.

• Letta and the cloud tier. A student’s memory is bounded by its blocks, not by how long they talk.
• Weaviate becomes LanceDB, on the device. The embeddings API becomes MiniLM in ONNX, shipped with the app.
• The TEU keeps three local tools. The calculator parses the expression instead of calling eval.
• The write-back blocks the reply again: the scheduler is there but never started. Measured, then named a design error.
• The model is a setting: point LLM_PROVIDER at a local Ollama and nothing leaves.

Two silent failures found on the way: BMJ promotions undone every turn by a second copy of the DLL, and vectors of two different widths between the course pipeline and the client.

Article EZFRICA/alu-education-agent

Implementation2026-09 → 10code + measurement

Akili, the current version

The tutor grows past the article: a guard in front of every exchange, whatever the modality, and a model chosen by measurement for each role.

• A topical guard, with a separate verdict for distress that is never counted and never stored.
• Text, voice and braille, all through the same guard.
• Kept out of memory on purpose: escalation events and the pupil’s notebook. The tutor never reads the notebook, as the UCCE agent never read its Inbox.
• The write-back is still awaited. In the background it measured 34.6% less perceived latency, but the decision is open.

EZFRICA/apu-akili

In progress2026-10not yet published

Urithi, handing on what a person knows

The prologue’s assistant for the elderly, turned around: the elder teaches. Urithi structures the knowledge during the exchange and publishes nothing until the guardian has heard it back and confirmed it.

• Evidence kept verbatim and attributed, and a library in Neo4j searched by meaning and by words, filtered by access level.
• Memory per knowledge: a block written on one knowledge never reaches an answer on another. Every block keeps its source.
• The learner reads what is remembered about them and deletes any of it, from every tier.
• Page-in: an L3 hit comes back into the DLL. A meter puts the write-back under a budget.
• Team memory, seen in TencentDB Agent Memory. Rejected: it would cross the two sides and the organisations kept apart.

Almost every idea here was already in UCCE in January: scoped memory, a source for each fact, memory the person can see.

How Urithi works, step by step EZFRICA/apu-urithi