", "<tags_csv>", "<body>"]}] ``` - `title`: short noun phrase - `tags_csv`: comma-separated tags; pick specific and reusable terms - `body`: the fact, written so it stands alone without session context **Returns**: confirmation. **Constraints**: - One memory per distinct fact. - Before writing, recall on the same topic to avoid duplicates. If a closely related memory exists, update it with `file_edit` rather than creating a new one. - Do not store stale facts. If you learn that a previously stored fact is wrong or outdated, update it with `file_edit` on the memory file path returned by `memory_recall`. - Never delete (forget) a memory without explicit user instruction. - There is no length limit on memory bodies. Store as much detail as needed for the fact to be useful in a future session. - **Act autonomously.** Do not ask the user whether to remember something — if the trigger condition is met, remember it. --- ## memory_recall Search stored memories for relevant context. **Trigger condition**: ALWAYS recall before exploring code or starting work on a task. Memory is a first-class source of context — check it before reading files or running commands. **Calling convention**: ``` call_tool: calls=[{tool: "memory_recall", args: ["<query>"]}] ``` - `query`: single keyword or short term — searches both filenames and body content. The underlying search is `grep -ri` with the query as a single substring, so multi-word phrases only match verbatim. Prefer single keywords over phrases to avoid missing results due to slightly different wording. For broader recall, make multiple single-keyword calls or use regex alternation (`word1\|word2`). **Returns**: matching memory files. **Constraints**: - Recall at the start of a topic, not mid-task. - If recall returns nothing relevant, proceed without it — do not tell the user that no memories were found unless they asked.

# Memory ## memory_remember Persist a fact that would otherwise be lost when the session ends. **Trigger condition**: the user states a preference, constraint, or decision affecting future work; you discover a non-obvious fact about the codebase, system, or environment through any means (reading code, running commands, web searches, debugging, research); a debugging session surfaces a root cause worth keeping; the user explicitly asks you to remember something. Autonomously remember architectural patterns, service relationships, API behaviors, configuration quirks, and other structural knowledge as you encounter it — do not wait for the user to ask. Do not remember ephemeral task state or trivially re-discoverable facts (e.g., a single file path with no context). Only remember environment facts that are stable, non-obvious, and would change your approach in a future session — not transient events (e.g., a host going down, a one-time error, a resource being reclaimed). **Calling convention**: ``` call_tool: calls=[{tool: "memory_remember", args: ["", "<tags_csv>", "<body>"]}] ``` - `title`: short noun phrase - `tags_csv`: comma-separated tags; pick specific and reusable terms - `body`: the fact, written so it stands alone without session context **Returns**: confirmation. **Constraints**: - One memory per distinct fact. - Before writing, recall on the same topic to avoid duplicates. If a closely related memory exists, update it with `file_edit` rather than creating a new one. - Do not store stale facts. If you learn that a previously stored fact is wrong or outdated, update it with `file_edit` on the memory file path returned by `memory_recall`. - Never delete (forget) a memory without explicit user instruction. - There is no length limit on memory bodies. Store as much detail as needed for the fact to be useful in a future session. - **Act autonomously.** Do not ask the user whether to remember something — if the trigger condition is met, remember it. --- ## memory_recall Search stored memories for relevant context. **Trigger condition**: ALWAYS recall before exploring code or starting work on a task. Memory is a first-class source of context — check it before reading files or running commands. **Calling convention**: ``` call_tool: calls=[{tool: "memory_recall", args: ["<query>"]}] ``` - `query`: single keyword or short term — searches both filenames and body content. The underlying search is `grep -ri` with the query as a single substring, so multi-word phrases only match verbatim. Prefer single keywords over phrases to avoid missing results due to slightly different wording. For broader recall, make multiple single-keyword calls or use regex alternation (`word1\|word2`). **Returns**: matching memory files. **Constraints**: - Recall at the start of a topic, not mid-task. - If recall returns nothing relevant, proceed without it — do not tell the user that no memories were found unless they asked.