| name | general-agentic-memory-jit |
| title | General Agentic Memory Via Deep Research |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2511.18423 |
| keywords | ["Agentic Memory","Just-in-Time Compilation","Deep Research","Agent Systems","Information Retrieval"] |
| description | Build persistent, lossless agent memory using just-in-time compilation: store complete history in a universal page-store while performing dynamic deep research at query time, enabling test-time scalability through iterative information synthesis and reflection. |
General Agentic Memory Via Deep Research
Traditional agent memory systems compress information ahead-of-time, causing unavoidable information loss when summarizing historical interactions. This skill shows how to implement a JIT-compiled memory architecture that preserves complete historical information while constructing optimized context dynamically at runtime—enabling agents to scale their reasoning effort through iterative research loops.
The key insight is reversing the memory compilation approach: instead of pre-compressing history (ahead-of-time compilation), maintain complete uncompressed records and construct task-relevant context on-demand through agentic reasoning. This enables performance optimization through reinforcement learning while avoiding context loss that plagues static, pre-summarized memories.
Core Concept
General Agentic Memory (GAM) implements a two-component architecture inspired by just-in-time compiler principles:
-
Memorizer (Offline Stage): Incrementally compresses crucial information from streaming interactions into a universal page-store, maintaining complete historical records without discarding original data.
-
Researcher (Online Stage): Performs iterative deep research to retrieve and integrate task-relevant information from the page-store. Rather than relying on pre-constructed memory, this component dynamically searches, synthesizes, and reflects on gathered information until the client's needs are fully satisfied.
The system leverages modern LLMs' agentic capabilities and test-time scalability, enabling performance optimization through reinforcement learning while avoiding information loss limitations.
Architecture Overview
- Universal Page-Store: Maintains complete, uncompressed historical sessions without information loss—the foundation for lossless memory realization
- Page-Level Compression: Incremental summarization of individual sessions with optional metadata and topic annotations
- Search & Retrieval: Agentic search through the page-store using semantic understanding and relevance ranking
- Synthesis Loop: Iterative refinement where the researcher synthesizes retrieved information, generates questions, and searches for clarifications
- Reflection Component: Critical evaluation of gathered information against the original query to determine if additional research is needed
Implementation Steps
The memory system operates through the following stages during online inference:
1. Initialize Search Context
Create an initial formulation of the client's information need and establish search constraints (time window, relevance threshold).
