简体中文 · English

U-Mem: Autonomous knowledge acquisition and memory management

Fidelity: Implements cost-aware tiered acquisition, semantic retrieval, Thompson sampling, and memory validation. The mini-suite does not include the paper's real LLMs, HotpotQA, or AIME inference.

Paper information

Field	Details
Paper	Towards Autonomous Memory Agents (arXiv 2602.22406)
Institution	National University of Singapore
First public release	2026-02-25
Author code	Anonymous review repository
Local adapter or method key	u-mem
Local reproduction code	src/auto_research/agent_research/

Original paper summary

Background and main changes

Traditional agent memory passively stores and retrieves information without deciding where to look when current knowledge is insufficient. U-Mem uses an acquisition cascade with rising cost: try self/teacher first, then research with tools, and finally consult an expert. Retrieval combines semantic similarity and Thompson sampling, and memory is validated and curated before write-back.

flowchart LR
    Q["Task + current memory"] --> C["Confidence check"]
    C --> S["Self / teacher"]
    S --> V{"Validated?"}
    V -- "No" --> T["Tool research"]
    T --> V
    V -- "Still fails" --> X["Expert"]
    V -- "Yes" --> M["Curated memory"]
    M --> A["Answer / action"]

Key figure from the paper

Figure 2: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations

Memory selection balances semantic relevance and exploration under uncertainty:

\[ s(m\mid q)=\lambda\,\mathrm{sim}(q,m) +(1-\lambda)\,\tilde p_m,\qquad \tilde p_m\sim\mathrm{Beta}(\alpha_m,\beta_m). \]

Success and failure update the Beta posterior. The cascade weighs the extra knowledge gained against call cost.

Offline and online results

The paper reports a 14.6-point improvement with Qwen2.5-7B on HotpotQA and 7.33 points with Gemini-2.5-Flash on AIME25. No production online A/B test is reported.

Local reproduction

evomem-mini covers four quadrants: within-episode versus cross-episode, crossed with knowledge versus execution. The implementation records each retrieval, validation, escalation, and write, escalating from a local knowledge source to tool research at low confidence.

auto-research agent-eval --method u-mem \
  --benchmark evomem-mini --episodes 120 --seed 42

Metric	Long-context	U-Mem
Joint success	1.0000	1.0000
Average cost	64.5000	3.0500
Final memory size	Full history	12

Stable metrics: agent-mini-suites-seed42.json.

Scope

The mini-suite's knowledge sources, experts, and validators are deterministic. It has no real search APIs, model hallucination, or token cost. It verifies the active-acquisition state machine and cost accounting, not HotpotQA or AIME scores.