The Question
Modern AI systems have become increasingly good at accessing organizational information.
But access is not the same thing as memory.
An organization's knowledge is distributed across documents, databases, systems, events and people. It changes over time. Sources conflict. Decisions are revised. Facts can become obsolete.
SARP asks what architecture is required for an AI system to maintain a persistent, evolving and inspectable representation of organizational knowledge.
Retrieval is not memory.
The Thesis
Organizational memory should not be modeled as a collection of retrieved text chunks.
It should preserve the lifecycle of knowledge:
- where information came from,
- what was observed,
- what was inferred,
- what was verified,
- what changed,
- what remains uncertain,
- and what evidence supports the current belief.
Generated conclusions should not silently become permanent truth.
Reference Architecture
- Sources
- Observations
- Evidence
- Facts
- Entities + Relationships
- Organizational Memory
- Reasoning
- Insights / Actions
What I'm Exploring
Observations vs facts
How should a system distinguish between something reported by a source and something accepted as organizational knowledge?
Evidence & provenance
How can every important claim remain traceable to the evidence and transformations that produced it?
Entity relationships
How should organizational entities, events and relationships be represented beyond document similarity?
Temporal knowledge
How should the system represent what was true, when it was true and how knowledge changed?
Contradiction
How should conflicting evidence be preserved rather than prematurely collapsed into a single answer?
Confidence
How should uncertainty be represented without turning arbitrary model confidence into truth?
Memory lifecycle
How should knowledge move through states such as observed, inferred, verified and superseded?
Human oversight
How should people verify, reject, correct and promote machine-generated knowledge?
Current Design Principles
- Evidence is a first-class object.
- Observations and facts are different things.
- Generated conclusions do not automatically become truth.
- Knowledge must remain connected to provenance.
- Time is part of memory.
- Contradiction should be preserved when unresolved.
- Humans participate in knowledge formation.
- Memory should improve agent control, not only retrieval.
These principles may evolve over time. This is active research, not a finalized doctrine.
Current Status
ACTIVE BUILD
The core organizational-memory architecture and its enterprise use cases are currently being developed and tested.
Current work includes evidence-grounded memory, temporal knowledge representation, entity relationships, human verification and production use cases.