How to use the Latent Briefing KV Memory agent skill

What an agent skill is and what this one does

An agent skill is a folder of instructions that an AI agent loads when a request matches the skill's description. The skill tells the agent how to approach a specific kind of task.

This skill covers hierarchical multi-agent systems where a coordinating model delegates to worker models. Instead of passing the orchestrator's full reasoning history into each worker prompt, the approach compacts that history inside the worker model's KV cache and keeps 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.

• Explains token explosion in recursive or hierarchical agents
• Describes Attention Matching (AM) KV cache compaction as the core engine
• Lists three inference-time changes: task-guided queries, a shared token mask, and MAD thresholding
• Helps decide between Latent Briefing, prefix caching, summarization, structured notes and retrieval

Install the latent-briefing skill

Download the ZIP for skills/latent-briefing from the directory page. Unzip it so the latent-briefing folder (containing SKILL.md and references/attention-matching-formulation.md) sits in one of these locations:

For a personal install, place it in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/latent-briefing/
• Project: .claude/skills/latent-briefing/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with natural requests

You do not need a special command. The skill loads when your request matches its description. The description lists phrases such as "share memory between agents", "KV cache compaction for multi-agent", "orchestrator worker context", "latent briefing", "reduce worker tokens", and "cross-agent memory without summarization". It also triggers on discussion of Attention Matching compaction, recursive language models with workers, or token explosion in hierarchical agents.

Plain questions work too. Name the problem you have, such as worker prompts growing with every call, and say what you want to compare.

• "How can my orchestrator share its memory with worker agents without summarizing it?"
• "Our workers' token costs explode in a recursive agent. Is latent briefing viable?"
• "Explain Attention Matching compaction and how it applies to orchestrator-worker systems."
• "Compare KV cache compaction with RAG for passing state to a worker."

What the skill covers in practice

When active, the agent explains the reasoning behind representation-level sharing, works through the trade-offs, and checks whether your infrastructure can support it. The skill notes that the main design question is whether the system controls the worker inference runtime closely enough to inspect or transform KV state.

The skill also gives tuning guidance. Higher tau in the median + tau * MAD threshold is more aggressive. Reported long-document QA results suggest lighter compaction for longer documents and more aggressive compaction for harder questions, but the skill labels these as hypotheses to re-measure on your own workload.

• Use task-guided query vectors derived from the current worker prompt
• Aggregate scores into one shared global mask rather than per-head subsets
• Use a robust threshold such as median + tau * MAD
• Re-measure worker tokens, total tokens, compaction overhead and accuracy on your target workload

Requirements and what you cannot do with it

The skill itself is guidance and conceptual explanation. It is not an installable tool. The technique it describes needs access to the worker model's inference runtime and internal KV cache. The skill states that API-only stacks, where internal KV tensors are inaccessible, are a poor fit, and it points to other skills (context-compression, memory-systems, multi-agent-patterns) for those cases.

The skill also assumes the orchestrator's trajectory can be represented in the worker's model space. If orchestrator and worker differ materially in tokenizer, architecture or attention layout, direct representation sharing may not be viable.

• Access to the worker model's inference runtime and internal KV cache, if you plan to implement the technique
• Runs scripts: no
• Uses network: no
• Not suitable for simple chatbots or single-agent apps

Safety notes

According to the directory assessment, this skill runs no scripts and makes no network calls. It consists of SKILL.md and one reference file. Review the files yourself after download, as you would for any skill you install.

The reported benchmark figures come from a public write-up on long-document QA workloads. The skill itself says to treat them as workload-specific evidence, not a general guarantee.

• No scripts are run by the skill
• No network use is declared
• Treat reported savings as workload-specific

Tips for getting useful output

Give the agent your architecture: whether you run your own inference, which models act as orchestrator and worker, and whether you can access KV cache. Those facts decide whether the skill's advice applies.

Ask for a decision, not just an explanation. The skill's decision framework helps you pick between prefix caching, structured notes, retrieval and Latent Briefing based on your bottleneck.

• Name your bottleneck: worker token cost, latency, or document access
• State whether you control the worker's inference runtime
• Ask for a measurement plan on your own workload