MaiBot Docs

English

简体中文

English

简体中文

Appearance

View and Manage Memory ​

Updated: 09/14/2026Word count: 897 wordsReading time: 3 min

MaiBot stores what it learns from chats in long-term memory, just like human memory. The 长期记忆 (Long-term Memory) page (/resource/knowledge-base) under the "麦麦资源管理" (MaiBot Resource Management) sidebar group centralizes memory management: query, import, correct, delete, and tune — all in one place.

Long-term memory

Memory Overview ​

Open the Long-term Memory page; the top tab bar is organized by purpose:

  • 记忆查询 (Memory query) - search memory content
  • 图谱 (Graph) - entity relation graph and evidence view
  • 审计时间线 (Audit timeline) - review memory changes per chat flow
  • 情景记忆 (Episodic memory) - view and rebuild episodic memories
  • 人物画像 (Person profiles) - query and maintain person profiles
  • 导入 (Import) - create and manage import tasks
  • 记忆检修 (Memory inspection) - maintain memory state and correct content
  • 删除 (Delete) - bulk delete and history rollback
  • 纠错历史 (Correction history) - view feedback and rollbacks

Search Memory ​

In 记忆查询 (Memory query), enter keywords (e.g. "game", "food") and filter by time or by user to see memories from a period or from chats with a specific person.

Memory query

Knowledge Graph ​

The 图谱 (Graph) tab shows relations between concepts like a mind map:

  • Each node is a concept (e.g. "Genshin")
  • Edges represent relations (e.g. "Genshin-game")
  • Click a node for details

The standalone 长期记忆图谱 (Long-term Memory Graph) page (/resource/knowledge-graph) provides full-screen visualization:

Long-term memory graph

Audit Timeline ​

Review memory changes for each chat flow:

Audit timeline

  • Memory audit events (add, update, delete, etc.) paginated in reverse chronological order
  • Filter by chat flow and event type
  • Change summaries are merged directly into the event list

Import Memory ​

The 导入 (Import) tab lets you teach MaiBot new knowledge manually:

Import memory

  1. Choose an import kind: 资料导入 (material import: text, file, or folder), LPMM OpenIE, or LPMM conversion
  2. Paste text or upload files
  3. Optionally set common and advanced parameters in the "导入参数" (Import parameters) dialog
  4. Start the import; the task list shows progress in real time

Correct Memory ​

When a profile or relation is inaccurate, correct it via 记忆检修 → 内容修正 (Memory inspection → Content correction):

  1. Set or remove manual overrides in person profiles
  2. Adjust nodes, relations, or weights in the knowledge graph
  3. Use feedback correction, delete-and-restore, or re-import to handle outdated content

Plain paragraph text currently has no arbitrary text editing entry; to correct it, delete the wrong source and re-import, or use the feedback correction mechanism.

Delete Memory ​

Don't want to remember something? The 删除 (Delete) tab supports:

Delete memory

  • Single delete: find the memory and click "删除" (Delete)
  • Bulk delete: select multiple items and delete together
  • Delete by source: delete all memories of a chat flow

⚠️ Note: deleted items go to the recycle bin and can be restored

The 纠错历史 (Correction history) tab shows feedback and rollback records:

Correction history

Person Profiles ​

MaiBot builds a "profile" for every user:

  • Personality traits (outgoing, introverted, etc.)
  • Interests and hobbies (games, anime, etc.)
  • Chatting habits (sticker usage, speaking style, etc.)

In the 人物画像 (Person profiles) tab or the 人物信息管理 (Person Info Management) page (/resource/person) you can:

Person info management

  • View profiles
  • Correct inaccurate descriptions
  • Add notes for friends

Retrieval Tuning ​

If MaiBot's memory is poor, run a tuning task to optimize retrieval (记忆检修 → 检索调优, Memory inspection → Retrieval tuning):

Retrieval tuning

  • The page keeps only the description and the start button
  • Tuning parameters live in the "调优参数" (Tuning parameters) dialog; they only affect the next tuning task, and defaults are usually fine
  • After a task completes, review the evaluation result and apply the recommendation with one click if it passes validation

Runtime Maintenance ​

记忆检修 → 状态维护 (Memory inspection → State maintenance) provides runtime self-checks, the auto-save switch, vector rebuild, paragraph vector backfill, import tasks, and delete operation records. The "更多操作" (More actions) menu in the top-right corner centralizes memory runtime status (including vector rebuild and data refresh).

State maintenance

Usage Recommendations ​

Daily Maintenance ​

  • Review memory regularly and delete useless content
  • Correct errors as soon as they are found
  • Manually reinforce important information

Improve Effectiveness ​

  • Teach the bot domain knowledge to make it smarter
  • Refine person profiles for more considerate conversations
  • Set memory capacity appropriately to balance performance and effect

Verification & Troubleshooting ​

Verify: import a piece of text; after the task completes, the content should be searchable in 记忆查询 (Memory query).

Import tasks stuck in queue or failing?

  • Confirm a working model provider is configured (import needs models for extraction and vectorization)
  • Check the failure reason in the task details

Memory query finds nothing?

  • Confirm the import task completed (status "已完成")
  • Check whether vectors are built; rebuild them in "状态维护" (State maintenance) if necessary

How long are memories kept?

Kept long-term by default. Memory evolution gradually decays old relation weights, and low-weight content may be marked for pruning; the exact behavior is controlled by A_Memorix's memory evolution configuration.

Do memories leak privacy?

Memory data is stored in local directories by default. Generating summaries, profiles, corrections, or vectors may call the model services you configured; confirm the data boundary according to your deployment and model provider.