Claude Code: Deterministic Recall on Every Turn

What gets wired

Claude Code is one of neuron’s two deterministic adapters — recall is enforced by the harness itself, not requested by agent judgment. neuron init wires four lifecycle points into Claude Code’s own hook system:

• SessionStart — a large, one-time payload (the architecture blueprint card, if Architecture Scan is enabled, plus an initial recall pull).
• UserPromptSubmit — a small, deduplicated pull on every turn, so a 50-turn session doesn’t re-inject the same entry 50 times.
• PreCompact — clears neuron’s per-session injection ledger when context compacts, so an entry injected early and then compacted away is eligible to surface again rather than being treated as already delivered.
• PreToolUse (on Bash calls) — the pre-command hook, firing a pre-execution memory lookup automatically on every shell command your agent runs, as additionalContext next to the tool result.

Where it writes

Hooks are written to .claude/settings.json — project-committed, project-local (gitignored), or user-global, whichever scope you choose during neuron init. The file is parsed and edited in place, only the relevant array elements touched, never wholesale regenerated. Claude Code also deduplicates hooks by command/args, so re-running init doesn’t double-register the same hook.

MCP config

neuron init also writes .mcp.json, so neuron mcp’s three tools (neuron_remember, neuron_recall, neuron_query_exec) are available to any Claude Code session even though the deterministic hooks above already cover recall and pre-command lookup without it.

Write-side compliance

Claude Code also gets the Stop hook, which forces one more turn with a reminder if nothing has been recorded to memory by the time your agent tries to end its turn — empirically verified to force a continuation rather than just suggest one.

Source: ADR 0014 — Recall Adapter Architecture.