Adkid-Zephyr/Liltloom
Language weaving: Chinese-first, user-controllable AI writing style memory layer, allowing AI to learn your expressions and call them when needed
Liltloom is a Chinese-first, user-owned writing-style memory layer for AI. It learns structured, reviewable language preferences from eligible user-authored text (filtering out tool results, code, logs and secrets), then compiles a bounded Style Context Packet only when you explicitly ask the AI to write in your style. Base mode stores aggregate features and rules rather than full conversation transcripts; every rule can be inspected, edited, locked, suppressed, deleted, and exported as versioned JSON. Deep Style (raw exemplar retention) and small-model analysis are separate opt-in features. Ships a native DSH settings page, /liltloom slash commands, the style_context model tool (off by default), and a portable core usable by other host adapters.
Install
dsh plugin --profile web add "$PWD"Source install per README (bilingual, English section verified): git clone https://github.com/Adkid-Zephyr/Liltloom.git && cd Liltloom && pnpm install && pnpm run check, then dsh plugin --profile web add "$PWD", then dsh --profile web --dump-config and dsh --profile web. Packed alternative: pnpm pack then dsh plugin --profile web add /absolute/path/to/liltloom-0.2.0.tgz. Current adapter targets DSH 0.1.0-rc.5 and is build-tested against 0.1.0-rc.6.
Compatibility
DeepSeek Harness Web profile with a storage-domain provider. Node ^22.19.0 || >=24, pnpm. Uninstall note: dsh plugin --profile web remove liltloom does not delete your data — run /liltloom clear confirm first.
Details
- Repo: Adkid-Zephyr/Liltloom
- Category: Memory Enhancement
- Stars: 4
- Version: Source/tarball 0.2.0 (no npm package)
- Last push: 2026-08-14
- First seen: 2026-08-14
Recent updates
v0.2.0: silent style learning (statistical + structured rules), explicit style-context activation, rule CRUD + lock/suppress, versioned JSON export, opt-in Deep Style and small-model analysis, DSH settings UI + /liltloom commands.
FAQ
- Does it store my conversations?
- Base mode stores aggregate features and rules, not full transcripts. Raw exemplars are opt-in (Deep Style) with strict filtering of secrets and bounded quotas.
- How do I make the AI write in my style?
- Activate explicitly: /liltloom use (next response), /liltloom use session (until /liltloom off), or the style_context model tool after /liltloom tool on.
- Can I take my data with me?
- Yes — /liltloom export writes readable versioned JSON; import/import-replace restores it on another machine.
Alternatives
Aik358/dsh-auto-memory · csyangwen/dsh-memory-evolve · Scorp1o117/dsh-tdai-memory