02.3: Server Memories

此内容尚不支持你的语言。

Server memories provide persona-scoped long-term memories stored for the active server or direct message channel.

Flow and ownership

The contributor buildServerMemoryContextItem in src/utils/text/context/memories.ts loads and formats memories for the current persona lineage:

Precondition check: Returns null if tomoriState.server_memories is empty or undefined. In nativeBuilder.ts, the contributor is skipped entirely during user impersonation.
Database retrieval: Queries serverMemoryRepository.loadServerMemoriesScoped(serverId, personaLineageId) for all memory rows scoped to the server and persona lineage. Missing scoping IDs or a failed load yield no rows. If row preparation fails, the block is omitted because cached content strings lack the channel scope and decay metadata needed to filter them.
Channel scope filtering: When channelMemoryEnabled is enabled, a memory with a non-empty channel_disc_ids is filtered out unless it contains the current channel ID or, inside a thread, the parent channel ID (channelScopeAdmits in src/utils/discord/channelScope.ts). An empty scope applies in every channel. A memory whose legacy #name scope from migration 111 is still unresolved (channel_scope_pending, read from legacy_channel_scope_tags) has an empty scope, so it applies in every channel until the backfill or an admin narrows it.
Decay filtering: Visible rows enter the shared recall collector. Dormant rows are withheld from this block and may return beside documents in the uncached tail; active rows retain their existing formatting. See the decay lifecycle.
Formatting and emission: Surviving rows are formatted as ID:${id} ${content} using formatMemoryWithId and wrapped in a section header:
- Guild channels: ## {botName}'s Memories about {serverName}
- Direct messages: ## {botName}'s Memories about this conversation with User

The block is emitted as one system-role item tagged KNOWLEDGE_SERVER_MEMORIES after passing through convertMentions.

The item also carries individual plain-text memory entries with internal identities and recency. NovelAI fits and renders those entries without tool IDs or tag prefixes; ordinary providers retain the formatted block. Placeholder and mention resolution applies to both projections.

Constraints and rationale

Stable memory IDs: Prompt IDs map to server_memory_id, allowing update_long_term_memory to target a row for replacement or deletion. Empty replacement content requests deletion.
Legacy tags are inert: Server memory keyword and #channel tags neither hide a memory nor appear in its prompt line. Lorebooks own keyword activation, and channel ids own channel scope.
Scopes are IDs: A stored scope survives channel renames. Document retrieval applies the same channel-or-parent rule in SQL (retrieveRelevantDocumentChunks). Data export leaves scopes out, so imported memories arrive unscoped.
Preset reassembly: Tagged as KNOWLEDGE_SERVER_MEMORIES, which SillyTavern preset reassembly flushes at the first knowledge anchor.

Source pointers

src/utils/text/context/memories.ts: buildServerMemoryContextItem.
src/utils/db/repositories/ServerMemoryRepository.ts: loadServerMemoriesScoped.
src/utils/memory/memoryId.ts: formatMemoryWithId.
src/utils/text/context/mentionNormalizer.ts: convertMentions.
Long-Term Memory Pipeline: Creation, modification, and deletion lifecycles.