# Memory as a Service (MaaS): Purpose-Bound Memory Mediation for Cooperative Agents

arXiv: https://arxiv.org/abs/2506.22815 · HTML: https://mlsanity.com/paper/2506.22815/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Agent memory systems, one of 16 directions in LLM agents and tool use.
- First posted on arXiv 28 Jun 2025.

## Placed in

- LLM agents and tool use / [Agent memory systems](https://mlsanity.com/agents/agent-memory-systems/)

## Abstract (arXiv)

Agentic programming is code-centered, while its useful memory context extends beyond code. A programming agent may draw on memory from test, review, build, and release agents; design, product, security, operations, and compliance agents; meeting, finance, calendar, and workflow agents; personal agents; and agents acting for other people. These memories can help agents optimize, debug, test, and evaluate software, while carrying different owners, purposes, recipients, and disclosure boundaries. We propose Memory as a Service (MaaS) as purpose-bound memory mediation: each invocation is evaluated by owner, requester, recipient, task, and declared purpose, and the mediator chooses whether to withhold, abstract, or reveal each candidate item. We formalize this by separating cooperative utility, disclosure leakage, and purpose-bound authorization, then ground the position with diagnostic stress tests on MAGPIE. Relevance-based retrieval reaches AUROC 0.570 and leaks 53.0% of private items; contextual-integrity prompting reduces leakage by 21.8 percentage points while leaving 32.6% residual leakage; and 4.5% of private items contain explicit safe-hint abstractions. These probes motivate memory governance as a separate design problem for cooperative programming agents.