Latent Briefing KV Memory
Shares orchestrator reasoning with worker models via compacted KV cache instead of long text handoffs.
- Stars
- 18,282
- Licence
- MIT
- Files
- 2 files
What it does
This skill guides an AI agent through a technique for multi-agent systems where a coordinating model hands work to smaller worker models. Instead of pasting the coordinator's long reasoning history into each worker's prompt, the approach compacts that history inside the worker model's internal memory (its KV cache), keeping only the parts relevant to the worker's current task. The skill explains the idea, its trade-offs against summarization and retrieval, and when it applies.
Great for
- Engineers building orchestrator-worker or recursive agent systems that suffer from high token costs
- Researchers studying KV cache compaction and Attention Matching for inference-time memory sharing
- Teams deciding whether representation-level state sharing is viable for their infrastructure
Skip it if
- API-only stacks where internal KV tensors are inaccessible, because the technique cannot run without access to the worker's cache (the skill itself says other skills fit better there)
- Simple chatbots or single-agent apps, where the token explosion problem does not arise
- Non-technical users looking for a ready-made tool; the file is guidance and conceptual explanation rather than an installable product
Try asking
- “Help me design an orchestrator-worker setup that stops re-sending the full trajectory to every worker call”
- “Explain how Latent Briefing differs from summarization and RAG for passing state between agents”
- “Evaluate whether my self-hosted worker model can support KV cache compaction for multi-agent handoffs”
- “Interpret benchmark results showing worker-token savings versus accuracy and compaction overhead”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 2 original files, the MIT licence and an attribution note.
latent-briefing.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip latent-briefing.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip latent-briefing.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Ask for a feasibility check first: confirm your runtime exposes KV tensors before planning an implementation
- Treat the reported savings as workload-specific evidence and benchmark on your own long-document tasks
- Tune the tau threshold empirically, since higher values compact more aggressively and the right setting depends on task and document length
What's inside
- references/attention-matching-formulation.md
- SKILL.md
Redistributed unmodified from muratcankoylan/Agent-Skills-for-Context-Engineering under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.