Whats missing? Suggest next feature

[vadash]

I am happy with v17. What do u guys wanna see next? This plugin will stay as auto as possible.

[grandfso]

Hi! I've been testing v17 with GLM-5 models via OpenRouter and found a few issues:

1. Prefill "none" crashes graph, reflections, and communities

When extractionPrefill is set to "none" (empty string), buildGraphExtractionPrompt, buildUnifiedReflectionPrompt, buildCommunitySummaryPrompt, and buildGlobalSynthesisPrompt all throw "prefill is required". The event builder handles this fine with prefill ?? '<think>\n' but the other builders don't have the fallback. Result: events extract successfully but you get 0 entities, 0 reflections, 0 communities — and the errors are swallowed silently.

2. LLM whitespace in character names creates duplicates

LLMs occasionally return names with trailing spaces (e.g. "Alice " vs "Alice"). This bypasses resolveCharacterName (lowercases but doesn't trim) and creates phantom character state entries. The Zod schemas for characters_involved, witnesses, entity names, and relationship endpoints don't trim either. You end up with duplicate characters in the filter dropdown and inflated character counts.

3. Extraction is blind to conversational commitments

The extraction rules and all few-shot examples are action-oriented (discovery, combat, intimate scenes). Conversational information like decisions ("I want you to check in daily"), promises ("I'll be there Saturday"), stated preferences ("I love hiking"), and ongoing rules gets rated low importance or skipped entirely. This is a problem for roleplay continuity — when a new chat session starts, the character forgets commitments and preferences that were explicitly discussed.

Related: momentary scene actions (a goodbye kiss, a one-time reaction, a compliment) get extracted with the same weight as durable information (rules, schedules, preferences). A durability lens — "would a character need to remember this next week?" — would help prioritize what actually matters across sessions.

4. Feature requests: time awareness and memory decay

• Time awareness: Memories currently have no sense of real-world time. If a character says "see you next Saturday" or "I'll be away for Easter", there's no mechanism to relate that to actual dates. Time-aware memories could enable characters to reference upcoming events, know when something was "a week ago" vs "yesterday", and recognize when a scheduled event has passed.

• Memory decay/expiration: Some memories are inherently temporary — "go do X tonight", "I'll be back in an hour". These become misleading noise after they expire. An optional TTL or expiration mechanism (either extracted by the LLM or inferred from temporal language) would let stale memories fade rather than persist forever with the same weight as durable facts.

[vadash]

1 and 2 is easy fix, 3 and 4 would require some testing (2 04-04 files)
https://github.com/vadash/openvault/tree/beta/docs/designs

Beta branch is updated. U can switch directly in ST extensions

[vadash]

I also added hints to UI

To fully use it need to clear memories

[grandfso]

I won't have opportunity to actually do tests until after the weekend, but reading through the time-awareness-transient-decay doc I have one comment - the defaults params for decay might be good or might not - I can't tell without testing - however I imagine different roleplay styles might have differenttime progression characteristics per message, and having sane defaults is good, but this could benefit from params exposed in the ui/advanced section?

edit: btw, is this BMAD? :)

[vadash]

edit: btw, is this BMAD? :)

its super powers modded. brainstorm and TDD with kimi k2.5/glm5.1, implementation with ralph loop glm 4.7. I use alibaba $10 and zai $6 sub

[CitrateSynthase]

I think this extension has a very strong foundation and great ideas behind it. It may, however, have an Achilles' heel at its present state.

The autonomous implementation may be there by design, after all it means less time for manual memory management and more time for the actual RP, but the fundamental problem is that LLMs can hallucinate, and with this extension, there is limited opportunity for the user to audit the output. It's possible to edit incorrect/imprecise memories, but not the character/entity entries, their relationships via graph edges, or community summaries that the LLM has come up with. Even if the hallucination rate is less than 5%, it can easily snowball.

One specific situation that I could not remedy was when the main LLM came up with some nonsense response right at the moment when OpenVault hit the token threshold for the next extraction, and OpenVault incorporated that nonsense into entities and community summaries. One possible way to prevent that would be to never include the last AI response into the extraction prompt, as a hallucination/swipe protection.

Another thing I've noticed is that sometimes it misses some characters and entities that are referred to by different names. For example, you can have an character that is first introduced as a "masked figure", and the character's name is revealed further down the line. The system may consider them as two entirely different characters, with no way to merge them. This also applies to other entities that are often referred to by different names/aliases. Some way to combine those entries that refer to the same thing would be great.

It may be a lot of work and it may go against your design philosophy, but I believe that giving the user all the tools to properly maintain the vault would really improve the experience and long-term usability of the extension, and may even make it a one-stop solution for both memory and complex worldbuilding.

Appreciate all your work, it's been a pleasure to follow this repo.

[vadash]

https://github.com/vadash/openvault/tree/beta

Version History

• 21.00 - Provides a manual mechanism to merge two entities
• 20.00 - Entity editing, deletion, and alias management with tab restructure from "World" to "Entities" + "Communities"
• 19.00 - Exclude the last N complete turns from extraction batches so hallucinated/swiped AI responses dont enter the memory