Paper summary: General Agentic Memory Via Deep Research

Problem

Most agent memory is static: it’s summarised ahead of time so it’s ready to use. According to the abstract, this inevitably loses information that turns out to matter later.

Method

General agentic memory (GAM) follows a “just-in-time compilation” principle: keep simple but useful memory offline, and build an optimised context at runtime for each request (Figure 1).

• Memorizer: writes a lightweight memory that highlights key history, while keeping the complete history in a universal page-store.
• Researcher: when a request comes in, it searches the page-store, guided by the light memory, and integrates what it finds, much like a deep-research agent.

The design benefits from frontier LLMs’ agentic abilities and test-time scaling, and can be optimised end to end with RL.

Key results

The abstract reports substantial improvements over existing memory systems in memory-grounded task completion.

Limitations

There’s no dedicated limitations section in the extracted text. Runtime research costs more per request than looking up a pre-built summary.