Skip to content

Cursor: Best-Effort Recall at Session Start

Cursor is neuron’s other best-effort adapter, occupying the same session-start-only fidelity cell as GitHub Copilot CLI — but the two aren’t equivalent underneath. Cursor documents fail-open behavior on its hooks, with an opt-in failClosed setting; Copilot CLI leaves failure, timeout, payload cap, and disable switch all undocumented. Cursor’s sessionStart hook also does not run in cloud or background agent execution modes — an execution mode Cursor itself offers — and user-level hooks (~/.cursor/hooks.json) never run in cloud agents at all, since there’s no home directory there.

Hooks are written to .cursor/hooks.json — either ~/.cursor/hooks.json (user-level) or <repo>/.cursor/hooks.json (project-level), matching the scope you choose during neuron init.

neuron init writes .cursor/mcp.json, giving Cursor direct access to neuron_remember, neuron_recall, and neuron_query_exec via MCP.

Cursor gets stop’s followup_message, which auto-submits a continuation with a reminder if nothing has been recorded to memory by the end of a turn, per Cursor’s own documented followup_message semantics.

Like Copilot CLI, Cursor has no per-turn hook point — a memory recorded mid-session won’t surface again until the next session starts. Cursor’s beforeShellExecution hook is also documented as permission/gating-only, with no context-carrying field, so the pre-command pre-execution lookup available on Claude Code and Codex CLI has no equivalent here either; this is a structural limit of Cursor’s own hook model, not an unresearched gap. Cursor keeps the neuron exec -- <command> CLI wrapper as its permanent path for pre-execution lookups.

Source: ADR 0014 — Recall Adapter Architecture, 2026-08-10 amendment.