NemoLore

NemoLore is a modular SillyTavern extension for long-running roleplay continuity. It maintains per-chat memories and summaries, retrieves relevant context, manages generated lore safely, and can run memory, summary, and lore helpers after replies.

This release-candidate branch uses the modular runtime exclusively. Legacy settings and data are migrated, but the retired legacy runtime is not loaded.

Requirements

• A current SillyTavern stable or staging checkout.
• Node.js 20 or newer to run the test suite.
• A working SillyTavern text-generation connection, or an OpenAI-compatible chat-completions endpoint for helper jobs.
• A compatible source configured in SillyTavern Vector Storage if semantic retrieval is enabled. Lexical retrieval remains available without it.

Installation

Clone the branch into SillyTavern's third-party extensions directory. The installed directory must be named NemoLore because SillyTavern resolves extension assets from that path.

Set-Location <SillyTavern>\public\scripts\extensions\third-party
git clone --branch dev/preset-architecture https://github.com/NemoVonNirgend/NemoLore.git NemoLore

Start SillyTavern, enable NemoLore under Extensions, and reload the page. Open the NemoLore drawer and confirm that a story profile and the management buttons are visible.

For an existing installation:

Set-Location <SillyTavern>\public\scripts\extensions\third-party\NemoLore
git fetch origin
git switch dev/preset-architecture
git pull --ff-only

Back up SillyTavern user data before changing branches. Do not install two copies of NemoLore under different folder names.

Story profiles and engine modes

NemoLore exposes four profiles:

• Short RP uses a large recent window with sparse helper activity.
• Long Form is the recommended balanced profile for stories lasting hundreds of messages.
• Episodic promotes scenes and consequences sooner.
• Epic favors precise provenance, hierarchical continuity, and aggressive extraction for very long chats.

Advanced changes become explicit overrides on the selected profile. Profile changes persist through SillyTavern's extension settings.

summaryEngineMode and loreEngineMode are always normalized to modular in this architecture. summaryContextPrecedence is always new-only. Old legacy values trigger migration; they do not reactivate the retired automatic workflows. When NemoTavern is present, live ownership checks prevent its Nemo memory, summary, or lore workers from running at the same time as NemoLore's modular owner.

Provider and helper-agent setup

The provider registry always includes:

• sillytavern: uses SillyTavern's active text-generation connection through generateRaw.
• async: available when Enable OpenAI-compatible provider is enabled and a complete endpoint is configured.

For an OpenAI-compatible provider, enter the full chat-completions URL, API key, and model, for example a URL ending in /v1/chat/completions. Credentials stay in SillyTavern extension settings; avoid sharing exported settings files containing keys.

Helper routing fields accept the provider names sillytavern and async:

1. Enable helper agents.
2. Choose a shared helper provider, or set memory, summary, and lore overrides.
3. Optionally choose a fallback provider.
4. Enable the desired post-reply workflows and adjust cadence in Advanced settings.

Jobs have bounded concurrency, timeouts, retries, circuit breaking, and replay deduplication. A helper failure is isolated from foreground chat generation. Failed jobs release their dedupe key so a later reply can retry them.

Context injection

Before foreground generation, NemoLore composes accepted preferences, the active modular summary, and relevant memory into the configured extension-prompt position. The inspector shows selected and omitted contributions and token use.

When a modular summary exists and the active profile enables message hiding, only messages outside the profile's recent window are excluded. This exclusion belongs to the modular summary policy; no legacy interceptor is invoked.

Managers

Use the buttons at the bottom of the NemoLore drawer:

• Manage Memories searches and filters the active chat's records, edits fields, reviews contradictions, archives/restores entries, and promotes important records to core memory.
• Manage Summary & Lore edits or regenerates the current summary, exposes lineage/source ranges, protects manual lore entries, previews generated operations, selectively approves changes, and merges duplicates.
• Open NemoLore Inspector shows context contributions, active memory/summary/lore state, helper jobs, provider routes, NemoTavern ownership, semantic-index status, and a guarded semantic rebuild action.

Memory, summaries, lorebook associations, and migration state are scoped to the active chat and survive chat switches and reloads.

Reviewed preference memory

Preference memory is opt-in and cross-chat. Candidate preferences never enter model context until accepted. Accepted records can be global or scoped to the current SillyTavern user persona. Rejected, disabled, candidate, and wrong-persona records are excluded.

See Preference memory for the data model.

Semantic retrieval

NemoLore reuses SillyTavern Vector Storage's selected source, model, credentials, and alternate endpoint. Configure Vector Storage first, then enable semantic retrieval in the NemoLore profile or Advanced settings. The inspector reports the inherited source, indexed/pending counts, and any error.

If Vector Storage has no compatible source, NemoLore reports semantic retrieval as unavailable and continues with deterministic lexical, entity, importance, recency, and type scoring. The in-process webllm and koboldcpp sources are not currently exposed to third-party extensions.

Migration

On first load, pre-profile settings are classified into the closest story profile. NemoLore records the prior policy in presetMigration, imports legacy summaries into modular stores, and links the historical NemoLore settings namespace to the canonical nemolore object. Migration is idempotent and source-preserving.

See Migration notes before upgrading or downgrading.

Development and verification

npm test
npm run test:smoke

Release candidate 1.4.0-rc.1 was exercised against SillyTavern staging 1.18.0 at commit 380e31e8c58d196969b6a0da74f431ba999c7e0a. See the changelog for verification scope.

Troubleshooting

• Drawer does not appear: confirm the folder is exactly NemoLore, the extension is enabled, and manifest.json loads bootstrap.js; then inspect the browser console.
• Helpers do not run: confirm helper agents and the individual workflow are enabled, the minimum-message/lore-signal rule is met, and the provider name is sillytavern or async.
• OpenAI-compatible requests fail: use the full chat-completions URL and confirm the endpoint accepts Bearer authentication and the configured model.
• No semantic matches: configure a compatible SillyTavern Vector Storage source or rely on the automatic lexical fallback.
• Wrong chat data appears: stop testing and capture the inspector snapshot and console. Chat switches are guarded; stale async work should be discarded rather than committed.
• Two automatic systems appear active: disable duplicate extension copies. NemoLore coordinates with supported NemoTavern host APIs, but cannot arbitrate unrelated forks that do not expose ownership state.

More detail is available in Troubleshooting, Configuration, Management, and Architecture.