Blog
Long notes on memory infrastructure.
Posts are shown in English when available. Missing translations fall back to the original.
AI Is Moving From Scientific Assistant to Scientist
Reading Google I/O 2026's Gemini for Science announcement as a shift in how science is conducted, not just an upgrade to scientific tooling.
Read →
Forgetting in General-Purpose AI Is Rational, But Not Rational for Every Domain
Automatic forgetting in general-purpose AI is a rational design, but tasks that require long-term accumulation need a different memory model.
Read →
If Druckenmiller Used an LLM, Single-Domain Questions Would Throw Away 90% of the Answer
The real power of LLMs is not single-domain answers, but question design that reveals the empty space between domains.
Read →
Prompt Engineering Is Over — Now It Is Answer Reading Engineering
In the LLM era, the next advantage comes from reading patterns in the distribution of answers, not just asking better questions.
Read →
The Place Where Lost Things Gather: Norfolk by Nunchi AI
Norfolk is a place where thoughts, notes, and conversations you thought were lost gather again as memory.
Read →
Even If Models Change How They Think, How a Company Remembers Matters More
As latent reasoning gets stronger, the value of verifiable external memory rises, making Nunchi AI's three business lines clearer rather than weaker.
Read →
The Two Philosophies of Agent Memory — Hindsight's Epistemology, Synapsis Engine's Operational Principles
Hindsight centers on the epistemology of memory, while Synapsis Engine focuses on operational principles. We compare and contrast the two architectures side by side.
Read →
96.6% Recall, 0% Portability
MemPalace proves that retrieval works. What it does not solve, and what the industry still lacks, is structured, portable memory that survives tools and runtimes.
Read →
Norfolk Is Becoming More Than a Notes App
Norfolk started as a notes app. It is now becoming a memory surface where humans and agents can both write, recall, and share context through the same infrastructure.
Read →
The Nunchi AI Memory Ecosystem
A four-layer architecture for portable agent memory: Synapsis for atomization, AMCP as the protocol, Nexus/Norfolk/MaaS as backends, and 3122 as the reference client.
Read →
Your Harness Will Change. Your Memory Shouldn't.
Memory lock-in is not solved by embedding memory deeper into the harness. It is solved by separating harness logic from portable memory and connecting them with an open protocol.
Read →
3 A.M.: The Night My Product Died and Came Back
Anthropic's Managed Agents announcement looked like the thing that would kill my product. By morning, it had clarified why memory should outlive models.
Read →
Linguistic Diversity and Emergence
A hypothesis that emergence in large language models comes not from scale alone, but from the density of intersections created by linguistic diversity.
Read →
The Performance Gap Between Agents Comes From Team Memory, Not the Model
Put Karpathy's personal knowledge-base workflow next to Stripe Minions and the pattern becomes clear: agent performance is driven by shared context and memory infrastructure.
Read →
Nobody Reads Docs Anymore So Who Is Your Onboarding For
AHOP proposes a missing onboarding layer between llms.txt, OpenAPI, and the real human-to-agent handoff.
Read →
Why Our 83.2% LongMemEval Score Matters More Than 99%
What our 83.2% LongMemEval result taught us about honest benchmarking, production memory systems, and where the real gains still are.
Read →
Completing the Agent Stack: AMCP and Memory Continuity
Why tools and collaboration are not enough, and why agents need a memory layer that survives sessions, runtimes, and clients.
Read →
Claude Channels Makes the Agent Reachable. Nexus Makes It Remember.
A narrow verified proof that Claude Code Channels handles live reachability while Nexus preserves memory across restarts.
Read →
Completing the Agent Stack with Memory Continuity
Positioning AMCP as the missing memory layer beside tools and agent collaboration.
Read →
Why Agents Become Unstable as Conversations Get Longer
A plain explanation of why large context windows still fail to behave like reliable memory.
Read →
Why Do Agents Lose Memory When Sessions Change?
A brief and clear explanation of the background behind AMCP's release and the agent memory continuity problem.
Read →