Explicitly store a long-term memory (takes effect immediately, enters recall). Call immediately when the user states an explicit preference/fact/decision/correction (e.g. "以后X", "我喜欢X", "记住X", "不要X") — do not wait for wrap-up. Keep content concise, self-contained, and independently understandable. Types: fact / preference / correction / sop / todo_context / event.

Feed recent conversation messages to the engine for automatic memory extraction (LLM or rule mode). Call it yourself at conversation wrap-up or when enough dialogue has accumulated — do not wait for a host hook. Extracted items default to pending (anti-poisoning), confirm via memory_pending + memory_confirm. Falls back to rule mode (zero external calls) when LLM is not configured.

Search long-term memories by keywords (keyword + embedding hybrid; falls back to keyword when embedding is unavailable). Returns matching entries with type, importance, and similarity. Call at the start of a new session/task (query = current topic) to restore context; also call whenever the user references past work, project history, or personal preferences.

List automatically extracted but unconfirmed items (anti-poisoning: only enter recall after user confirmation). Two kinds:

1. Pending memories (atom): handle with memory_confirm / memory_reject

2. Pending corrections: handle with correction_confirm / correction_reject Call after memory_extract / memory_capture to surface what is waiting for user confirmation, and remind the user.

Confirm a pending memory (enters recall; correction/preference/sop types also refresh the user persona). Call when the user approves an item listed by memory_pending.

Reject and delete a pending memory (incorrect extraction or poisoned content).

Read the L3 user persona markdown (stable summary of user preferences/behavior rules). Call at the start of a new session to load the user profile and preferences. Returns the merged view by default (global base persona + current project overrides, with per-line scope).

Manually save/overwrite persona markdown. Defaults to the current project layer (overrides that project part of the merged view); scope=global writes the global base persona. For maintaining the persona or project-specific behavior rules.

Read recent hot-scene summaries (timing signal for proactivity: what is being worked on recently, how hot).

View memory system statistics (atom count, type distribution, pending items, persona status).

Evaluate whether a conversation excerpt deserves proactive suggestions (correction / followup / automation / skill / todo). Call when the conversation shows signals worth reminding: an unfinished todo, a repeated mistake the user corrected, a follow-up promised, or an automation opportunity. Core principle: silence is also a skill. At most 1 per call, session budget limits, none during do-not-disturb hours. Returns newly created suggestions (may be empty). trigger: session_end (default) / session_mid (realtime, strong signals only, max 1) / manual.

List suggestion records, filterable by status (suggested / accepted / ignored / never).

Accept a suggestion. For memory_correction types, writes a behavior correction and refreshes the persona (accept = explicit approval). For automation/todo types, tries to actually create it (requires host-injected executor), otherwise returns executable instructions. Pass host to label the current host (e.g. claude-code / kimi) for fallback wording.

Ignore a suggestion (counts toward frequency learning: similar suggestions converge in weight; repeated ignores auto-silence that type).

List unified ActionCards (cross-source action inbox). Current source is suggestion engine; future agent / automation / bridge sources land here. Card status uses unified semantics: pending / accepted / dismissed / resolved.

Get a single ActionCard by id with full fields (source / priority / target / privacy / duplicateKey / evidence).

Confirm a pending behavior correction (from corrections.json, usually extracted by memory_extract rule mode). On confirmation: writes a correction memory and refreshes the persona. ID comes from memory_pending correction entries.

Reject a pending behavior correction (incorrect extraction or poisoned content). ID comes from memory_pending correction entries.

Generate a daily review guide: organize today's work into long-term memories and generate improvement suggestions. The returned template guides you through memory_recall / memory_extract / scene_summary / suggest_now. Suitable at session end or on a daily schedule.

Cold-start guide: teaches this session how to use ProactiveAgent memory & suggestion tools (when to use memory_capture/recall/extract, how to confirm pending memories, restraint principle). Call once on first mount in a new environment/project.