AdminChat Memory and Conversation Settings

Memory stores the current user's durable working preferences, project conventions, decisions, and references. It is not chat history, knowledge-base retrieval, Workspace files, or SystemMessage. It never grants tool permissions or overrides current user instructions and system safety rules.

Open the Memory button in the AdminChat sidebar, or choose Memory → Memory / conversation settings in the launcher or current conversation. Nearby question marks open this guide or its relevant section. Chinese UI locales use Chinese documentation; all other UI locales use English. See also Chat basics and Workspace access.

Getting started

1. Choose a Memory mode in the new-session launcher, or change it in the current conversation settings.
2. Open Memory settings and manually add a non-sensitive working preference under Entries, such as “Prefer concise bullet lists.”
3. Select your global scope or the current Workspace and save the entry. A session without a Workspace cannot edit Workspace memory.
4. Send a relevant question in Read only or Learn mode and inspect Loaded this turn.
5. For automatic learning, select Learn, explicitly say “From now on, remember…”, and check the system update receipt and entry list.

An assistant saying “I remembered it” is not proof of persistence. Extraction can finish after the assistant's reply. Only the system's confirmation of a successful database write establishes that the update was saved.

Off, Read only, and Learn

Mode	Uses durable entries this turn	Automatic learning
Off (default)	Does not inject Memory entries	No
Read only	Loads relevant entries	No
Learn	Loads relevant entries	Attempts extraction, validation, and saving when eligible

The mode is stored per session. The launcher configures the session about to be created, not every existing session. Turning Memory off does not delete entries or disable independently configured SystemMessage values. Read only restricts automatic learning; it does not prevent manual edits in settings.

Learn requires an available PromptRange Memory prompt provider and Chat model. Read only does not initialize learning prompts or call the extraction model. Learning uses an additional model call and may incur costs; not every conversation produces an entry. The current implementation looks for explicit English or Chinese triggers such as “remember,” “prefer,” “correction,” or “forget” in the current user's message. Overlong inputs and recognized requests not to save are skipped. Implicit preferences may be missed; add them manually when necessary.

User-global scope, Workspace scope, and storage

• The isolation boundary is tenant + current user. “My global” is not shared memory for all administrators.
• Workspace scope comes from the current authorized session. Settings show your global entries and your entries for that Workspace. Without a selected session, you can manage your global entries.
• When the same internal fact key exists globally and in the current Workspace, retrieval prefers the Workspace version. Similar text does not necessarily mean identical internal keys.
• The primary storage is an encrypted database record. The system does not automatically create or read a Workspace file named Memory.md. Such a file is an ordinary file, and Workspace sharing does not automatically expose database memory.
• Clearing chat history, turning off a mode, and deleting memory are different operations.

Editing, pinning, archiving, and deleting entries

Use Entries to add or edit individual records:

Field	Meaning
Scope	My global or the current session's Workspace
Tier core	A standing candidate, still subject to the context budget
Tier topic	Selected by lexical relevance to the question; a topic is required
Tier archive	Retained but not loaded; this is not deletion
Kind	preference, project, decision, or reference
Pinned	AI cannot modify or delete it; manual editing remains possible, and loading is not guaranteed
Expiry	Excluded from loading after expiry; not automatically deleted

Current limits are 80 entries across all scopes per tenant/user, 300 characters per entry, and 40 characters per topic. Entries in other Workspaces still count toward the total. At capacity, additions are rejected rather than evicting another Workspace's entries.

History retains at most 100 changes. Deleting an entry removes its historical content and evidence while retaining deletion metadata; it does not delete source chats, downloaded files, or backups. Conflicts preserve the newer database version. Keep a copy of your draft and reopen settings to reconcile it instead of blindly overwriting. Unsaved Markdown changes block single-entry writes to avoid conflicts.

Loaded-this-turn reports and update receipts

The report shows entries provided to the model in the latest turn on this page, with selection reasons. It does not prove that the model used them and is not a list of all stored entries. Relevant Topic entries and Core candidates are selected; archived, expired, and other-Workspace entries are excluded. Entries that exceed the budget are skipped whole rather than truncating facts.

The current Memory context budget is 1000 tokens, with at most 300 tokens for Core entries. Recognized models use the corresponding tokenizer; unknown models are explicitly reported as a conservative UTF-8 estimate, not an exact token count. The report includes the memory boundary instructions but excludes SystemMessage, conversation history, and model protocol overhead. It is not the complete request cost; consult actual model Usage for billing. Reports and update receipts are page-local and disappear on refresh.

Distinguish stored entries, entries loaded this turn, and entries updated this turn. Update receipts report global/Workspace change counts or reasons that no update occurred:

• Off/Read-only mode or an input that did not trigger learning;
• no candidates, no differences, or pinned entries;
• capacity or revision conflicts;
• invalid output/evidence, rejected safety checks, extraction timeout, or a failed call.

Extraction uses the current user's original message and a bounded selection of related existing memories. Assistant replies, attachments, tool output, and the whole transcript are not sources of new factual evidence. Do not infer persistence from an assistant's promise when validation fails. The turn's PromptCode can help diagnose extraction behavior.

Markdown preview, editing, and download

These tabs expose a Markdown projection of your global entries, not Workspace entries. Download exports the current editor contents, potentially including an unsaved draft; it is not a backup of all database memory.

Use this format, retaining the English heading spellings:

# Memory

## Core

- Prefer concise bullet lists.

## Topic: documentation

- Include a short example when explaining an API.

## Archive

- An old convention retained for reference.

Use single-line entries starting with - and keep one entry per fact. Arbitrary Markdown document structures are not supported. The editor accepts at most 32000 characters; entry and topic limits still apply. Saving Markdown treats new/modified entries as manually pinned entries and attempts to preserve the identity of existing entries with identical text.

Clear draft changes only the editor. Global entries are removed only after Save changes, without removing Workspace entries. Manage Workspace records individually under Entries.

SystemMessage versus Memory

• My SystemMessage affects all of the current user's Chat sessions, not only the selected session.
• Global SystemMessage is configurable only by a super administrator and affects all users in the current tenant, or the whole site in single-tenant mode.
• Each accepts at most 2000 characters. Saved changes apply from the next turn, independently of Memory mode.
• They apply to Chat conversations, not internal structured-generation tasks such as titles or Memory extraction. They never grant tool permissions or override safety/authorization rules.

Use Save changes for personal settings and the separate Save global SystemMessage button for tenant-wide settings. Changing a system prompt is not the same as adding an entry; disabling Memory does not remove these prompts.

Deployment, privacy, and troubleshooting

• After upgrading Admin, apply its migrations for the actual database provider and restart the appropriate service. Persist both the database and ASP.NET Core Data Protection keys; multiple instances need consistent key configuration.
• Legacy encrypted files are imported only when no database record exists. Existing database records, including cleared records, take precedence. Legacy files can remain as migration backups; apply retention rules and stop old instances from writing the old storage.
• If Learn is unavailable, check PromptRange and model configuration first. No automatic additions in Read only is expected.
• If Workspace entries are missing, confirm that you selected an authorized session associated with that Workspace.
• If the reply says “remembered” but no entry appears, inspect the system receipt, mode, trigger eligibility, pins, and capacity rather than assuming success.
• Do not enter credentials, passwords, sensitive personal data, or third-party secrets. Encryption and automatic validation are not complete sensitive-data detection guarantees. Downloaded Markdown is ordinary plaintext and must be protected separately.