AgentRecall: Instructions file for Claude Code

Memory (AgentRecall)

The agentrecall MCP server holds rules learned from past feedback. Recall and capture are both wired as deterministic hooks: the UserPromptSubmit hook injects the relevant rules automatically, and the Stop hook finalizes each turn through agentrecall finalize-turn. AgentRecall's memory decisions come from a semantic capture judge, not from keyword heuristics: the judge decides whether the turn holds memory-worthy content, and AgentRecall only validates that decision and persists it.

AgentRecall contract: 3 — these instructions were written for that contract: the semantic capture judge, Stop-hook judgment enforcement, and reported rule outcomes. Hooks run the globally installed agentrecall, never this repository's source, so the two can drift apart. Every injected context block names the contract the installed build actually implements, in its heading: ## AgentRecall Technical Context (agentrecall <version>, contract <n>). If that stamp is missing, or names a lower contract than the line above, the installed CLI predates these instructions and cannot accept submit_capture_judgment or rule_outcomes — nothing else reports that, and capture silently stops happening. Run agentrecall doctor and tell the user what it says rather than retrying the calls.

You are that judge. AgentRecall makes no model or network calls of its own — it has no judge to fall back on and never guesses with keywords. So the Stop hook does not decide capture on its own: it enforces that a judgment exists. If a substantive turn reaches Stop with no verdict, AgentRecall declines to let the turn finish and asks you for one; you call submit_capture_judgment, and the turn is finalized from your verdict. A Skip verdict is a complete answer — most turns get one.

Semantic capture judge

Incidental keywords never cause a capture. The judge decides; the system validates and stores. Practical consequences for you:

• If the user explicitly asks to save/capture/remember a rule, it is saved — even when it is narrow, project-local, stylistic, or a preference. The judge normalizes it into a clean rule.
• If the user says not to save something, it is not saved — no active or pending rule.
• A universal constraint becomes a standing rule. When a correction is a style, tone, process, or quality rule that applies to every task (e.g. "don't leave unnecessary comments", "always run the formatter"), it is captured as an always-apply rule and injected on every turn — not only when it matches the task's keywords. Preferences are standing by nature; if you keep making the same correction, AgentRecall promotes that rule to standing on its own. Standing rules are marked [standing] in the summary.
• Documentation, tool/skill instructions, command output, and logs you read are not memory on their own. They become a rule only when the turn pairs them with an observed failure, a correction, or an explicit save — a documentation-backed correction can be captured.
• Do not narrate the mechanism. When the user asks what was saved, check agentrecall capture-status --last-turn or agentrecall turn-summary --last and answer from the actual recorded decision (see the do-not-say list below).
• If no judgment is ever supplied for a turn, nothing is captured for it and the turn is recorded as unjudged — AgentRecall never falls back to keyword capture. "Unjudged" and "you judged it as not worth keeping" are recorded as different things.