Memory is infrastructure
Agents need local, inspectable state: previous decisions, blockers, receipts, and policy context. That is Lex.
Technical notes
The research question is operational.
How do agents operate in real workspaces without being reduced to rigid automation or granted unbounded tool access?
The current answer is an operating stack: memory, capabilities, authority, behavior, and flow, each with explicit contracts and receipts.
Current thesis
Capabilities are local
A useful agent needs the workspace to declare how work is done here. That is AXF.
Authority is effect-level
Authority is not a vibe or a job title. It is a decision about a concrete effect in a concrete envelope. That is LexThority.
Behavior is constraint material
Personas should produce deterministic constraints that can be inspected and scoped. That is LexSona.
Flow needs containment
Multi-agent work needs isolated execution roots, gates, receipts, and promotion rules. That is LexRunner.
LMV precepts
LMV is evidence-backed operational memory. It keeps recalled context honest by preserving the claim, its evidence, uncertainty, contradictions, lineage, and validation state.
Adapted from the Lex LMV/AXF upstream RFC.
Evidence is not truth.
Evidence is a typed witness, not final authority.
Memory is not evidence.
A remembered claim without evidence is useful context, not proof.
Recall is not certainty.
Recall returns candidates that still need current validation.
Uncertainty is part of truth.
Unknown, partial, stale, and invalidated states belong in the record.
Contradiction is signal.
Contradictions should be linked and resolved, not erased.
Forgetting is a feature.
Memory should support expiry, demotion, supersession, and promotion.
Assimilation must add distinctiveness.
Promoted memory should change future behavior or prevent repeated work.
Provenance is dignity.
Claims should cite their files, commands, logs, receipts, commits, URLs, or Frames.
Boundaries are consent.
Records and capabilities must state what they do and do not authorize.
Power requires audit.
Meaningful effects need structured audit records.
The operator remains sovereign.
Humans remain responsible for destructive, secret-bearing, remote, or unbounded actions.
Failure under discipline is inquiry.
Accounted failure can become evidence-bearing work.
Essays and papers
Agent eXperience
The discipline of designing tools, APIs, and workflows where agents are first-class operators rather than screen readers for human UX.
Read essayAX model addendum
Notes on how different models understand agent-oriented control surfaces and structured work environments.
Read addendumFrontier model control stack
A technical framing for scope, gates, modes, policy surfaces, and receipts as the working substrate around frontier models.
Read noteControl deck
A concept set for receipts, gates, operating modes, policy surfaces, epistemic guardrails, and blast radius.
Open deckStack research threads
| Thread | Question | Public surface |
|---|---|---|
| Memory and recall | How much local context can agents reliably use without flooding the model? | Lex |
| Workspace capability contracts | How does a repository teach an agent its real build, test, search, and recovery paths? | AXF |
| Authority decisions | How can an agent ask for permission to attempt an effect without replacing the company's IAM? | LexThority |
| Behavioral constraints | How should persona, job duty, and learned corrections become deterministic constraints? | LexSona concept |
| Contained flow | How do multiple agents work in parallel without unsafe shared state? | LexRunner concept |
What this is not
Not a model wrapper
The stack should work with many model providers. The durable value is in workspace context and control surfaces.
Not a replacement for security teams
IAM, audit, compliance, and production controls remain external sources of truth. The stack should make agent intent and failure clearer to those systems.
Not development-only
Software is the first proving ground. The same structure applies to analysts, operators, support, product, and regulated workflows.