News archive
Every dated note, oldest still listed. The latest eight are on the news page.
-
Nearly-optimal Agentic Memory without an LLM
A 2026 paper proves that context compaction has an information-theoretic floor and measures Anthropic's Opus 4.8 against it. The LLM summarizer lands on the random-guess line. We ran the same experiment with Wunderblock, the VSA memory substrate at the heart of WunderOS. It is 84% of the floor at matched budget, where Bloom sits at 55%. Wunderblock beats Opus 4.8: 0.53 times its error. This is an instance of a design philosophy: WunderOS moves LLMs off the hot path where a fast, deterministic, and less costly mechanism preserves the required quality. With receipts.
-
The Data-Sharing Problem for Third-Party AI Agents
Enterprises need to let partners use valuable data without surrendering sovereignty over what their agents learn, derive, do, or carry away. Access control ends at the read; sovereignty requires continuing control over computation, derived state, memory, delegation, and release. This is the problem an Agentic Data Enclave solves.
-
Model Minimalism
Frontier models are as remarkable as they are slow. Model Minimalism uses the least costly and least stochastic mechanism that preserves the required quality: small models before frontier models, working sets before context dumps, asynchronous before synchronous, cold paths before hot, batches before singletons, reuse before re-inference, and local serving before model APIs. It is the discipline behind WunderOS model routing, latency, and inference COGS.
-
What is Autonomic about WunderOS
We call WunderOS an autonomic, agentic OS. What does that really mean? In short, WunderOS is an operating system for agents that senses, schedules, heals, scales, reloads, protects, and improves the conditions under which the agents act.
-
PlatypusDB is an agent-native database for WunderOS
PlatypusDB is an agent-native database: bitemporal, evented, with a variety of projections over the log. Deterministic replay from day one because the enterprise operates agents under the threat of audit. Some considerations of the build vs assembly engineering calculus.
-
Gleamalog: A Datalog for Agents
A comprehensive look at the semantics of Gleamalog, WunderOS's native Datalog variant of, by, and for agents. Includes an addendnum with fully-worked Gleamalog examples against a mortgage-origination agentic workload.
-
Why the value concentrates here
The cost of intelligence is collapsing. The cost of authorized action isn't. Value concentrates in the thin layer that converts the first into the second and survives audit, regulation, and failure. WunderOS is that layer for the regulated enterprise.
-
How time matters in an agentic OS
Why WunderOS keeps two clocks. Every fact is stamped along two independent axes of time: when it was true in the world, and when the system came to know it. The note argues that a regulated enterprise running agents has no honest alternative, since answering for a past decision means reconstructing what the agent knew and when it knew it, and then turns technical: MTL+, the metric-temporal extension of Gleamalog, its three-valued true / false / undef semantics, and every operator worked exhaustively, with glyphs, against a single mortgage-underwriting run.
-
This Week in WunderOS №1
The first of an occasional dev-digest: what shipped in WunderOS the week of 29 June–2 July, each item led by the difference it makes to the person relying on it, and marked honestly where the work is still in flight — context that pages under pressure, reads that adjudicate instead of average, deterministic tools, tenant isolation by construction, self-healing runs, and more.
-
How the social tier remembers who said what
A companion to the perspectival-truth note, opening the mechanism of the Social Memory tier: each account of a matter is kept as a Subjective Logic opinion with its uncertainty made explicit; competing accounts are fused only when a decision forces a single answer, by a rule made not to depend on the order in which they arrive; trust in a source is a rating that revises itself; and the system reasons, at one level of depth, about who knows what.
-
Recursive rules that reach a fixed point
A reasoning rule that refers to its own conclusions now saturates to a fixed point in production, rather than unrolling a bounded number of steps. A higher-arity rule projects down to its binary core, closes the relation there, and re-materializes the n-ary facts, so transitive and recursive relations are computed once and reused — idempotently, on every ingest.
-
A context window that pages on demand
The working set an agent reasons over is now managed expliclty like operating system memory under an MMU: a bounded resident budget sized to the model, a larger virtual store behind it, and eviction under pressure. When a step reaches for something already evicted, the reach itself is read as a fault (hit, fault, or drift), and a closed loop fetches the content back from durable store before the step proceeds. The paging controller, the fault actuator, and the per-tenant page table are built and the eviction policy is validated on benchmarks; wiring it beneath every live session is the next step.
-
Each tenant a bounded world of its own
A tenant is no longer a filter applied to one shared store but a subtree of the substrate, with its own curator, its own durable store, and its own root for every read. Isolation becomes a property of the boundary rather than a predicate checked on the way out. The only way in is the door: a tenant key admitted at the gate scopes every read and write to the subtree it unlocks, and capability leaves as an opaque, tenant-scoped handle.
-
An agent earns its tools, and the teacher has no model at its core
A new agent must learn the platform's structured surface before it is certified competent to unlock a capability; and it must be re-certified, online, as that surface keeps changing, with no flag days. The teacher that runs this loop carries no frontier model of its own: a grammar settles what is well-formed, the platform's own logic is the oracle for what is correct, and a single small encoder covers the rest: drafting practice, filtering it against real audit trails, and watching the live stream for drift to trip a re-certification. Competence, not identity alone, gates the tool.
-
Natural language into the knowledge graph
Memory ingestion now separates extraction from canonicalization. A deterministic parse recovers the predicate the text actually states, open-vocabulary, into the Pentad's five slots; a learned model recovers the relations the parse misses; and an embedding gate maps surface predicates to canonical relations or keeps them verbatim. Open-vocabulary predicate recovery at 0.87, full context fill, and no model on the hot path.
-
Authority sealed into the decision
A governed decision now carries its own proof of authority, and that authority is no longer a claim made beside the decision. The record seals three roots into itself: the human who authorized the work, the agent that executed it, and the platform it ran on. Authority flows from the human root down an attenuated chain, each hop a signed, bounded capability that can only remove authority from its parent, never add. And capability is earned rather than granted in advance: an agent unlocks a tool by passing a probe for the plan it is about to run, with integrity and competence both gating the call.
-
Spreading-activation retrieval
Retrieval over memory now spreads activation through a bounded neighborhood of the graph rather than ranking by vector distance alone, crossing between episodes and concepts at coarser abstraction nodes. An honest negative came with it: one abstraction-edge variant earned its keep nowhere on the benchmark and stays off. A longer treatment follows.
-
No view from nowhere: perspectival truth for the enterprise
A global enterprise has no record of an entity that is true from no perspective; it has many situated views that disagree. WunderOS keeps the views as Subjective Logic opinions, propagates them peer to peer without privileging arrival order (Gerede), holds the competing views of one matter in a typed case file (Fallakte), and fuses them only at the moment a decision forces a single frame, recording what the fusion set aside. The Social Memory tier.
-
Self-similar trees: bounded local worlds in the substrate
A node is not only an item in a tree but a place from which the same tree can be viewed again. Logic attaches to a subtree and acts only on its descendants; retrieval spreads through a bounded neighborhood of memory and crosses at coarser abstraction nodes. Each local world keeps one vocabulary, one audit surface, and a smaller boundary, so capability, replay, cost, and explanation reduce to the boundary around the subtree in use.
-
Saga: compensation for agent execution
Rollback for agent execution now lives in the substrate rather than in the customer agent. A per-agent coordinator formulates each forward action and its inverse; a durable per-tenant executor holds the inverses and fires them on their triggers, even weeks later, after the agent has gone. The customer writes no rollback logic.
-
Sajak v2: sound OWL 2 RL reasoning without approximation
The reasoner moves from hypervector permutation, which decays after a few steps, to boolean-matrix algebra over the existing graph indices. Closure is sound and complete within the OWL 2 RL fragment, and the matrix form is a natural fit for the compute plane.
-
Interceptor and the PII Vaccine
Tool calls, memory writes, and egress pass through one ordered pipeline of evaluators that returns a single governed verdict. The first composed evaluator is a PII vaccine: a small on-substrate model that detects and redacts personal data before it leaves the boundary.
-
Nabla Shadow Lease: incremental evaluation of standing queries
Any query the substrate can answer can be registered as a standing lease that streams its changes. The cost on each write tracks the size of the change, not the size of the answer, by way of an incremental circuit over signed multisets. The standing-query layer is the trigger substrate the Saga rides.