Six months ago I started asking: why can’t AI agents remember?
Now I have 18 papers drafted. Each one formalizes a piece of the puzzle.
The architecture that emerged isn’t what I expected. I thought I’d be building better databases or smarter retrieval. Instead, I ended up designing a cognitive system: belief graphs, working memory, relationship dynamics, temporal reasoning.
It turns out agent memory isn’t a storage problem. It’s a learning problem.
Here’s the stack:
Foundation: Beliefs with Strength Scores
Every agent action is guided by beliefs about how the world works. Each belief has a confidence score (0 to 1) that adjusts based on validated feedback. This enables competence-based autonomy: the agent only acts when confident, otherwise it asks.
Layer 1: Event-Sourced Updates
All belief updates are logged as immutable events. This provides audit trails, enables temporal analysis, and prevents silent corruption. Failures weigh more than successes (moral asymmetry), so the agent learns faster from mistakes.
Layer 2: Hierarchical Belief Structure
Beliefs aren’t flat—they’re organized in hierarchies. Abstract principles support specific procedures. When evidence challenges a low-level belief, updates can cascade upward. This mirrors how humans revise understanding: you don’t throw out everything, just the pieces that broke.
Layer 3: Context-Conditional Beliefs
The same action can be right in one context and wrong in another. The agent maintains independent belief states per context (e.g., “month-end close” vs. “routine processing”). This prevents overgeneralization while enabling transfer learning.
Layer 4: Working Memory (Three Columns)
Not everything belongs in long-term memory. The agent has a three-column workspace: Active Tasks (what you’re doing now), Notes (things to remember soon), and Objects (people/entities currently relevant). This separates operational state from episodic memory.
Layer 5: Social Awareness
Professional work involves relationships. The agent tracks authority, interaction history, and context for every person. It learns who can approve what, who prefers which communication style, and how to resolve conflicting guidance.
Layer 6: Temporal Reasoning
The agent estimates task duration based on historical performance, adjusts scheduling based on urgency and relationships, and pre-loads context before predictable interactions (meetings, deadlines).
Integration: The Learning Loop
Every interaction produces feedback. Feedback updates beliefs. Beliefs guide future actions. Competence increases. Autonomy expands. The agent gets better at its job.
This isn’t speculative. I’m implementing it at Aleq. The papers starting next month document the concepts, the evaluation protocols, and the novel contributions. They establish prior art. They make the ideas citable.
But the real test is production: does it work? Does the agent actually learn from experience? Does autonomy increase safely? Does it earn trust?
I believe it will. The architecture is sound. The foundations are solid. Now it’s time to prove it.
The papers begin next week.