Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents¶
L1 core-mechanism diagnostic. Runs repeated probes, conflict rejection, and memory admission; does not connect to the GitHub Copilot SDK.
Paper information¶
| Field | Details |
|---|---|
| Paper | Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents |
| Organization | First-author affiliation: Microsoft |
| First public release | 2026-09-09 |
| Authors' code | No public author repository was found on 2026-09-14. |
| Adapter / method | grounded-memory |
| Local reproduction code | src/auto_research/agent_research/latest_20260914.py |
Original paper¶
Motivation and method¶
Give the asynchronous memory curator least-privilege, read-only tools to verify, scope, and refresh candidate memories before writing.
flowchart LR
I[Input and public state] --> M[Grounded Memory mechanism]
M --> A[Auditable intermediates]
A --> O[Output or update]Key figure from the paper¶
Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.
Core computation or operation¶
The local code implements the decisive operation as a deterministic reference kernel and explicitly reports gate, selection, projection, graph-edge, or state-update statistics. Tests cover formula invariants and boundary conditions.
Paper results¶
The paper reports CLBench pass rate rising from 39% to 73%, queries falling from 8.8 to 4.7, and task-agent cost falling from $3.38 to $1.68. These are paper results, not directly comparable with the scaled-down local run. No online gains are inferred in the absence of an online A/B test.
Local reproduction¶
See the three-seed metrics. All candidates share the public fixtures, budget, and seeds. The results verify actual execution of core state transitions only.
Reproduction boundaries¶
The local run checks repeated probes, conflict rejection, and memory admission without the GitHub Copilot SDK.