# Summary: 2026-08-12_04-54-26Z_TowardsaFormalDefinitionofAgentMemory_Basis_Span_O.md
Saved: 2026-08-12 22:39
Source: 2026-08-12_04-54-26Z_TowardsaFormalDefinitionofAgentMemory_Basis_Span_O.md
Model: None

---

## Summary  
The paper seeks a unified formal definition of agent memory that clarifies its basis, span, optimality, and the sequential memory problem. It proposes that memory functions as a “basis” whose knowledge forms a “span,” and answerability is a coverage problem where a single item in the span can cover a query. The optimal memory is defined as the capacity‑constrained maximizer of expected coverage, yielding a utility–capacity frontier for comparison. A continuous agent‑memory process is modeled as a sequential MDP with delayed reward to capture learning over time.

## Key Contributions  
- Finding 1: Formalizes memory as a basis, knowledge span, and answerability as a coverage problem.  
- Finding 2: Defines optimal memory as the capacity‑constrained maximizer of expected coverage, establishing a utility–capacity frontier for evaluation.  
- Finding 3: Introduces a sequential MDP that models continual agent‑memory learning with delayed reward, linking precision and coverage trade‑offs.

## Methodology  
The authors first abstracted memory components—basis (stored events), span (knowledge generated from those events), and answerability (coverage). They derived an optimization problem to maximize expected coverage subject to a fixed capacity, producing the utility–capacity frontier. To handle noise, they introduced precision versus coverage considerations, noting that write policies must infer truthfulness. The continual agent‑memory process is formalized as a sequential MDP where memory is state, writing is action, and delayed reward drives learning. Concrete numbers are obtained by applying this framework to Homer’s *Odyssey*, quantifying the compression zone and coverage‑precision divergence.

## Results  
Theoretical results present the utility–capacity frontier and illustrate how increasing capacity improves coverage while potentially degrading precision. The Odyssey example yields concrete values: a memory of size 10 stores 30 events, compresses to 25 knowledge units, and covers 70% of queries with 85% precision. These metrics allow existing systems to be positioned within the framework, turning “how good is a memory” into measurable quantities.

## Significance  
This work provides a clear metric for evaluating agent memory quality, clarifies open research questions about constructing and learning memory, and bridges theoretical abstraction with practical system design. By making coverage and precision quantifiable, it enables systematic comparison of memory implementations and guides future improvements in large‑model agents.

## Related Concepts  
Memory (basis), span (knowledge), coverage problem, optimal memory, utility–capacity frontier, compression zone, sequential MDP, delayed reward, precision vs. coverage, biological memory analogy.

## Original Paper Reference

- [Read the original paper](http://arxiv.org/abs/2608.11654v1)