MEMA vs. continual learning: three kinds of AI memory
Context windows, explicit MEMA records, and model-internal continual learning solve different problems. Here is the evidence boundary for MEMA, Nested Learning, and a planned HOPE integration.
Knowledge on AI engineering · CH · 2026
Structured, verifiable knowledge on AI engineering.
We're two engineers who can't let go of the question of reliable AI. Here we share what we learn, openly — notes, foundations and open research on building AI you can actually trust, instead of just using it.
§01 Topics
The craft behind reliable AI.
Five areas that decide whether AI becomes an advantage or a liability.
Data quality
High-quality, structured data, the fuel for everything else.
Model fluency
Know each model's strengths and weaknesses, and pick the right one.
Frameworks
Compose agents, skills, loops and workflows correctly.
Context engineering
Keep context structured and high-quality as it grows.
Deterministic layer
A verification layer for reliability, security and compliance.
§02 Latest articles
From the workshop
The AI literacy obligation: what Swiss companies must now train
Since 2 February 2025, Article 4 of the EU AI Act requires every provider and deployer to ensure a sufficient level of AI literacy among its staff. Swiss firms are in scope the moment their AI touches the EU market. Who the obligation hits, what it demands, and what appropriate measures look like.
Memory an AI agent can inspect
MEMA separates explicit records, retrieval, policy, audit, and integrity checks. Here is what those controls establish—and what they do not—at the current public and internal baselines.
Don't build a company brain. Scope memory small.
Errors in memory don't repeat, they breed. So the size of a memory is the size of its risk. The fix isn't one brain that knows everything; it's many small, isolated memories, each scoped to one cohesive use case.
§03 Research
mema, verifiable memory for AI agents.
Our open research into structured AI memory. The publicly inspectable baseline is package 2.22.13 at c37dbf7; the internally verified v2.26.0 baseline at d542fba is not public. Production readiness is not asserted.
§04 Authors
Who writes here.
Two engineers who would rather show how something works than sell it.