Abstract
An AI agent runs inside a host program that calls hooks at fixed points of its loop. A personal memory graph serves the agent through four of them. At session start it tells the agent who it is, what it was doing and what it meant to do. Before each prompt it adds the few records that are clearly relevant to the request. After each answer it takes the text as a short-term trace that becomes a memory only if it survives offline consolidation. Before each tool call it does two things. A gate extracts the names an edit changes, finds their readers outside the edited files up to the point where data leaves the program, adds questions drawn from the project's history of defects, refuses the call once, and lets the identical call through when it carries the agent's answer for each reader and question. A recall step searches the same graph with the tool call itself, so lessons about the agent's own habits come back at the step they concern rather than when a person repeats them. There is no separate store of lessons. The document gives the steps, variants and the closest prior work.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Prokhorov, Artem, "Memory at the Moment of Action: Agent-Host Hooks That Bring a Memory Graph to Each Step of an AI Agent, with a Gate That Names the Readers of an Edit Before It Is Made", Technical Disclosure Commons, (November 05, 2026)
https://www.tdcommons.org/dpubs_series/12770