A learning architecture built around experience.
Technical brief
Learning beyond the first interpretation.
Causal Trajectory Learning by Causality.
An extracted memory reflects what a system understood when it wrote it. CTL retains a representation of the experience behind that abstraction. Individual decisions form temporal trajectories, accessible at multiple resolutions. The system compares an unfolding trajectory with relevant prior experience and brings selected context into ongoing execution. Outcomes and feedback inform which experience is useful. The same history can be revisited for a different question, and operational experience remains separate from the underlying model's weights.
Request the technical briefDetailed implementation material is shared under NDA where appropriate.
The architecture
- Temporal and multi-resolution representation
- Experience-guided inference
- Adaptation to judgment and changing context
- Continuity beyond a single model
Go deeper
- Representation and retrieval design
- Runtime integration
- Evaluation and research collaboration
Research questions
The architectural questions are specific: does preserving trajectory structure improve transfer? Does decision-time retrieval help beyond task-start retrieval? Can the same experience support a new interpretation? Evaluate these separately, alongside quality, harmful transfer, and full operating cost.
Technical discussions can examine the representation, runtime behavior, and available evaluation artifacts. Comparative performance should be tied to a defined task, baseline, and measurement. See workflow examples.
Architecture
General capability. Accumulated experience.
Language models provide general capabilities. The runtime connects them to action. CTL represents what happens through that interaction and brings relevant experience back into the decision process.
People and environment
Context, constraints, interaction, and feedback.
Agent runtime
CausalityReasoning, action, and observed results.
Causal Trajectory Learning
CausalityTemporal experience, multi-resolution retrieval, and guidance during execution.
Foundation models
The underlying language and reasoning capabilities.
Representing experience
A decision belongs to a trajectory.
CTL begins at the scale of an agent's individual iteration, then connects those iterations through time. The representation supports access to a task, a sequence, or a decision point within its surrounding context.
- T0
Context and goal
What the agent was trying to achieve and the conditions it faced.
- T1
Reasoning
The agent's recorded approach and the considerations it expressed.
- T2
Action
The step it took in the environment.
- T3
Consequence
The observed response and resulting state.
- T4
Feedback
The judgment or correction attached to the experience.
One experience, new questions
The same experience can teach something new.
A difficult incident may reveal both a useful diagnostic approach and the standard of evidence your team expects before a change.
Those are different lessons from the same experience. CTL keeps that experience available to be revisited from the needs of the present task, so learning can continue beyond the first summary.
One retained incident
ALater question A
Which diagnostic step was useful?
Checkout errors rise after a release
Agent compares error logs with the release diff
Finds a changed payment timeout
Proposes a rollback
Engineer asks for a failing test first
Rollback approved after the test fails
BLater question B
What evidence did the engineer require?
Experience-guided inference
The current decision becomes the query.
A task description is only the beginning. As the agent reasons and acts, its developing trajectory provides a more specific basis for finding useful prior experience. CTL is designed to bring that experience into the process before the next action.
Recognize the situation
Use the developing task state, rather than only the original request.
Recover a relevant path
Find prior decision segments while retaining their surrounding context.
Consider what followed
Connect the earlier approach to its outcomes and feedback.
Guide the next step
Bring selected experience into ongoing reasoning and action.
Retain the new experience
The resulting trajectory becomes available for future use and reinterpretation.
Research direction
Starting with agents. Building toward adaptive intelligence.
Our research asks how experience can remain useful across changing tasks, models, and environments.
Software agents are the starting point. The broader direction is intelligence that learns through interaction, including future work connecting language, perception, and action.
We are interested in working with researchers and technology teams exploring the same frontier.
Discuss the researchTechnical questions
What changes with CTL?
Why retain experience instead of just extracting better lessons?
A lesson reflects what seemed important when it was written. A later problem may need a detail or relationship that the original interpretation overlooked. Retaining the connected experience allows the system to revisit it for a different purpose, while still using concise lessons where they help.
Why isn't storing and retrieving the full history enough?
History makes information available; it does not determine which part should guide the next action. CTL organizes experience across tasks, sequences, and individual decisions, then compares relevant portions with the agent's developing situation. Surrounding context, prior outcomes, and feedback help distinguish a useful path from a merely similar one.
What does causal mean in Causal Trajectory Learning?
CTL organizes experience around actions and their observed consequences, in context and over time. This lets the system compare earlier decision paths and what followed them. A recorded sequence does not itself establish causation; controlled evaluations are needed to isolate how a particular mechanism changes outcomes.
How do you distinguish useful learning from remembering or repeating?
Recall shows that an agent can recover earlier information. Useful learning shows up when that experience helps it handle a new situation appropriately, including adapting an earlier approach rather than replaying it. A meaningful evaluation includes cases where prior experience should help and cases where applying the same lesson would be wrong.