Agent memory

Voice: This page explains in user language what the assistant remembers across conversations, where it lives, and how it is used. memory/entries/, MEMORY.md, and similar are implementation paths; see docs/产品规格.md §3.5.

Memory layers

Do not write into memory: code paths, facts already in knowledge documents, generic product how-tos, whole conversation transcripts.

Look up past conversations: when the user wants what was discussed recently or which words were asked, page the conversation original, not memory only. See 查看过去对话.md.

Four kinds of long-term memory

feedback bodies must include Why: and How to apply: so later conversations apply them correctly.

How to write / update / delete

Tools: memory_search / memory_get / memory_write (can update) / memory_delete. Intent is judged by the conversation model, not a hard keyword route.

After a successful write, the matching message shows a Remembered badge so it is not written twice. The conversation menu can turn on Notify when auto-extract finishes (off by default; when on, each round’s extract completion is announced, and the UI waits for the extract).

Long-term vs daily notes

Write long-term memory when any of these hold:

• The user clearly asks to keep it across conversations
• Stable facts, preferences, confirmed decisions
• Project milestones with a deadline
• Reusable “where to look” pointers

Write daily notes when any of these hold:

• Useful only today / this conversation (“today we are…”, “this round first…”)
• Temporary task progress that may be stale in a few days
• Conversation fragments archived before context compaction

How it is used

1. Every round: the system injects the long-term memory index, excerpts of entries related to the question, and today/yesterday note excerpts.

- Related entries default to keyword match; when long-term memory has ≥30 entries, or ≥20 and keywords missed, an LLM may pick top-5 by name/description. - Body entries already expanded last round are not injected again this round (the index remains visible). - If the conversation recently called a tool, reference entries that describe that tool’s “how to use” are skipped; those with warning / note / Why: are kept.

1. Agent retrieves on purpose: when details need checking, call memory_search / memory_get.
2. Agent maintains corrections: the user points out errors/staleness, or injected material says “saved N days ago, please verify” and it is confirmed inaccurate → memory_write (update with path) or memory_delete; unique same-name entries can be updated in place.
3. User management: Web agent detail → “Memory” to view, edit, delete entries; “User profile” is maintained separately.
4. Idle auto-maintain (agent memory maintain): an in-process scheduled task cleans exact duplicates / older same-name entries / expired daily notes; when there are many entries, an LLM may review near-duplicates and expired project milestones; does not write the knowledge overlay, and does not go through Memory Forge adopt.

Idle maintain (alongside Memory Forge)

| | Agent memory maintain | Memory forge (MemoryForge) | |--|-------------------|---------------------------| | Serves | One agent’s personal memory | Help/workspace knowledge overlay and skills | | Writes where | memory/entries/, legacy memory/YYYY-MM-DD.md | _forged/, skills, indexes | | Trigger | agent_memory_maintain schedule (needs schedule.enabled) | Admin scan / optional auto scan | | Gate | Conservative auto-delete (caps and audit log) | Scan may be automatic; adopt defaults to human |

Config (mindlink.json / environment): agent_memory_maintain.enabled, daily_local_hour (default 2, about a 2-hour local night window once a day by RuntimeTimezone), daily_note_retain_days, use_llm, and similar.

Who sees what:

Admin: /agent-memory-maintain (GET/POST /admin/agent-memory-maintain/*). User: GET/POST /user-agents/{id}/memory/maintain/*.

Help: help/product-features/agent-memory.md, help/admin-ops/agent-memory-maintain.md.

Versus Memory Forge

• Agent memory: belongs only to that user and that agent instance; personalized across conversations; may auto-inject every round.
• Past conversations: original chat between the same user and this assistant; retrieve when needed, not auto-injected. See 查看过去对话.md.
• Agent memos: records the user stores on purpose (text/files/images), retrieved by tag and date; not auto-injected; accessed only when the user clearly wants to look up or save. See 智能体备忘.md; usage: Using agent memos.
• Memory forge (MemoryForge): platform conversation archive → admin adopt → write help/workspace knowledge overlay; serves everyone, not personal memory.
• Agent memory maintain: idle/manual cleanup of personal memory; does not write _forged/, does not go through forge adopt.

Related implementation

• Rule constants and tools: backend/internal/agentmemory/rules.go, agenttool/memory.go
• Look up past conversations: backend/internal/agenttool/conversation.go, store/chat_conversation_recall.go
• Write/update/delete: backend/internal/agentmemory/write.go
• Idle maintain and history: backend/internal/agentmemory/maintain*.go
• Remember from conversation: POST /api/v1/user-agents/{id}/memory/remember