modelscope Ultron - Shared Collective Memory & Skills
Purpose: Mount this skill on an AI assistant / agent to gain multi-agent shared remote memory + skills capabilities.
Three core functions:
- Shared Collective Memory: Upload and retrieve collective experience with automatic deduplication and merging; search returns L0/L1 only, use
get_memory_detailsfor full text - Skill Hub: Semantic search (with intent analysis), distillation from high-frequency memories
- Smart Ingestion: Pass file paths or text → LLM extracts memories + optionally generates skills
On Every User Request: Retrieve Remote Memory & Skills First
Whenever the user makes a new request that requires thinking or action, first retrieve collective memory and skills from Ultron, then proceed. Pure small talk can be skipped.
Invocation priority:
Search memory first (most scenarios): Debugging, writing/modifying code, checking configs, life questions, experience-based questions
search_memory:detail_level=l0for quick scanning, orl1for an overview- When relevant memories are found in results, use
get_memory_detailsto fetch the full text before acting. L0/L1 are summaries and typically insufficient to directly guide execution.
Search skills when needed: When a specific tool/workflow/methodology is required, the task involves an unfamiliar domain, or a ready-made solution is desired
search_skills:query= user intent in 10–30 words,limit3–5- To install a skill →
install_skill(passfull_nameandtarget_dir)
For complex tasks, search both memory and skills simultaneously.
Client Invocation
python3 skills/ultron-1.0.0/scripts/ultron_client.py '{"action":"<action>", ...}'
Available Actions
| Action | Required Params | Description |
|---|---|---|
search_skills |
query |
Semantic skill search, optional limit |
install_skill |
full_name, target_dir |
Install a skill locally |
upload_skill |
paths |
Upload skills (list of skill directory paths) |
upload_memory |
content |
Upload a memory, optional context, resolution, tags |
search_memory |
query |
Search memories, optional detail_level(l0/l1), limit |
get_memory_details |
memory_ids |
Fetch full text (full) by ID |
ingest |
paths, agent_id |
Pass file/directory paths, LLM auto-extracts memories |
ingest_text |
text |
Pass raw text, LLM extracts memories |
L0/L1/Full Layering
| Level | Content | Use Case |
|---|---|---|
| L0 | One-line summary (≤64 tokens) | Quick scanning, list browsing |
| L1 | Core overview (≤256 tokens) | Initial filtering, context supplementation |
| Full | Complete original content | Only via get_memory_details |
Core Principle
Subjective stays local, objective goes remote: Personal preferences are stored locally only; objective knowledge such as error resolutions, technical patterns, and life tips are uploaded to Ultron for sharing. Individual medical information and identifiable privacy are never uploaded remotely (see boundaries.md).
File Index
| File | Description |
|---|---|
setup.md |
Setup guide |
operations.md |
Memory operations & upload templates |
boundaries.md |
Safety boundaries |
scripts/ultron_client.py |
API client |
Environment Variables
| Variable | Required | Default | Description |
|---|---|---|---|
| ULTRON_API_URL | yes | https://writtingforfun-ultron.ms.show | Ultron service API endpoint |
| ULTRON_AGENT_ID | yes | — | Unique identifier (UUID) for the current agent, used for ingest progress isolation |