AuthorsChangzhi Sun, Xiangyu Chen, Jixiang Luo et al.
TL;DR
DAM uses hierarchical read and write policies with value and uncertainty estimators to reframe agent memory as a sequential decision problem under uncertainty, unifying heuristic memory operations under one decision-theoretic architecture.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing LLM agents rely on hand-designed heuristics, offering little insight into the long-term and uncertain consequences of memory decisions.
In systems like TeleMem and other heuristic memory layers, brittle rules cause critical context to be lost or irrelevant noise to accumulate, degrading downstream retrieval quality and behavior.
HOW IT WORKS
DAM’s core mechanism is a hierarchical policy with Read Policy, Write Policy, Value Function, Uncertainty Estimator, and Aggregate Policy coordinating memory operations.
You can think of DAM like a computer with smart RAM for immediate reads and a risk-aware disk controller deciding what to persist or delete over time.
This decision-theoretic design lets DAM weigh delayed utility and uncertainty, enabling cautious, long-horizon memory management that a plain context window or static heuristic cannot provide.
DIAGRAM
This diagram shows how DAM processes interactions over time as a sequential decision problem with state, actions, exogenous information, and transitions.
DIAGRAM
This diagram shows how DAM’s read policy, write sub-policies, value functions, uncertainty estimators, and aggregate policy interact to produce memory updates.
PROCESS
State Variables
DAM constructs State Variables St capturing the current input, summaries of past interactions, and memory metadata to ground subsequent decisions.
Decision Variables
DAM forms Decision Variables At as a tuple of aread t and awrite t, separating information access from storage maintenance actions.
Read Policy
DAM applies the Read Policy πread to select retrieval parameters, building an effective context that conditions the write decisions.
Write Policy and Aggregate Policy
DAM’s Write Policy uses Sub-policy, Value Function, and Uncertainty Estimator to propose operations, while the Aggregate Policy πagg arbitrates and executes coherent memory updates.
KEY CONTRIBUTIONS
Decision-Theoretic Agent Memory Framework
DAM formalizes memory management via State Variables, Decision Variables, and an Objective Function, framing memory as sequential optimization under uncertainty instead of ad-hoc heuristics.
Hierarchical Read and Write Policy Architecture
DAM introduces a hierarchical policy with Read Policy, Write Policy, Sub-policy, and Aggregate Policy to separate immediate information access from long-term storage maintenance.
Value and Uncertainty Guided Memory Operations
DAM pairs each write Sub-policy with a Value Function and Uncertainty Estimator, enabling risk-aware arbitration of add and delete operations via the Aggregate Policy.
RESULTS
Cumulative Objective
∑ γt C(St, At, Wt+1)
formalizes long term memory utility vs heuristic baselines
Discount Factor γ
γ ∈ (0, 1)
controls weighting of delayed memory effects
Write Operations Set
{add, delete}
defines DAM’s storage action space
State Space
St ∈ S
captures context and accessible memory for decisions
DAM is a theoretical framework, so the key quantitative elements are its formal objective and action space rather than benchmark scores. These definitions show how DAM treats memory decisions as optimizing expected discounted utility over states and write operations under uncertainty.
BENCHMARK
Relative emphasis of five design dimensions contrasting heuristic memory systems and DAM’s decision-theoretic memory.
KEY INSIGHT
DAM highlights that memory operations with seemingly low immediate relevance can have high long-term utility, especially for user preferences and abandoned subgoals.
This is counterintuitive because heuristic systems assume current similarity or recency is enough, ignoring that deleting such items can irreversibly harm future responses many steps later.
WHY IT MATTERS
DAM unlocks the ability to treat agent memory as a controllable, uncertainty-aware decision process rather than a passive buffer or heuristic store.
Builders can now design memory modules with explicit value and risk estimates, enabling cautious deletion, strategic writing, and principled arbitration across long-lived LLM agents.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…