Keep measured results separate from supplier estimates.
RememberedEvidence rule
RetrievedRelevant test
Later question
What changed in the revised cooling design?
Model parameters can also change through learning↗
Experience can be remembered, found again, and learned from.
Three different ways to carry work forward.
How it works
Three ways to build on experience.
Memory, retrieval, and changing model parameters solve different problems. Together, they let a system use earlier experience without treating every stored fact as a weight update.
Persistent memoryAcross sessions
First session
For Northstar, separate measured results from supplier estimates.
Project memoryKeep evidence categories distinct.
Later session
Compare the revised design with the last test.
Session 01Session 02
01 / Persistent memory
A new session does not have to start from zero.
Useful information can remain available after an interaction ends. When work resumes, Arkhe can carry forward relevant project context instead of asking for the same background again.
In the example, the rule about measured results and supplier estimates remains part of the Northstar project context.
RetrievalCurrent task / relevant context
Supplier notes
Earlier drawings
Cooling test / rev B
Brought into this task
Cooling test / revision B
Source retained with the passage
Question
What changed since the last test?
Available recordsSelected evidence
02 / Retrieval
Find the right context for this question.
Retrieval selects relevant information when it is needed. A prior test, document, or memory can enter the current task instead of forcing the model to rely on what fits in a single prompt. In this example, the selected record remains identifiable.
When the team asks about the revised cooling design, the earlier test is the record that matters.
Adjustable weightsA separate learning path
Before an update
One learned configuration
↗
After an update
A revised configuration
Parameters shape later responses. They are not a searchable copy of a conversation.
03 / Adjustable weights
The model can change, too.
Arkhe has adjustable model weights. As learning occurs, parameter updates can change how it responds to later inputs. That is a different mechanism from saving a note or retrieving a source.
The visual represents a qualitative change in parameters; it is not a plot of actual Arkhe weights.
What comes nextOne informed response
01 / Project memory
Separate measured results from supplier estimates.
02 / Retrieved test
Cooling test, revision B.
Draft response / illustrativeThe revised design shows a measured change in the last test. The supplier estimate is a separate input.
Context with boundariesAnswer with provenance
04 / Working together
A later answer has a history.
Project memory can preserve standing context. Retrieval can bring the most relevant record into view. Model learning can shape behavior over time. Each has a distinct role in the next answer.
The example is illustrative; real behavior depends on the configuration, available data, and learning controls.
One project, more than one encounter
Follow what carries forward.
The Northstar example is illustrative. It shows the roles of memory, retrieval, and adjustable weights without assuming a particular update schedule.
Separate measured results from supplier estimates.
Persistent memory
Measured results and supplier estimates stay distinct.
Retrieved record
Cooling test / revision B
Selected for the current task
MODEL PARAMETERSLearning path
First session
The project has a rule.
For the Northstar reliability review, the team asks Arkhe to separate measured results from supplier estimates. The project context can remain available after the conversation ends.
Context has boundaries
Learning with clear distinctions.
What a system remembers, what it retrieves, and what it learns in its parameters are separate paths. That distinction matters for technical clarity and for rights in the underlying data.
01
State has a boundary.
Remembered project context and customer-specific state remain distinguishable from general model learning.
02
A source stays a source.
Retrieved material can be identified as the context used for a response, rather than silently becoming a model claim.
03
Learning has its own path.
Adjustable weights do not mean every conversation or customer record automatically becomes training data.
Model and platform
From model capability to useful work.
ASIC Intelligence develops the foundation and the product surfaces that make it accessible.
01 / Model family
Arkhe
The model family at the center of ASIC Intelligence's work on persistent memory, retrieval, and adjustable weights.
A02 / Platform
Pipwick
An agent- and user-facing platform for building applications and workflows with intelligence capabilities.
