openclaw-knowledge-runtime（OpenClaw 知识 runtime）

OpenClaw Knowledge Runtime

What This Skill Does

Use this skill to design or implement a standalone knowledge runtime that can:

1. Read layered memory and knowledge sources.
2. Retrieve the most relevant knowledge for the current role, objective, and signals.
3. Link knowledge to entities, genes, tasks, and prior events.
4. Write stable findings back after successful runs.

Why Install It

This skill is useful when an agent already has memories, logs, tasks, and reusable assets, but they are still scattered across unrelated files or stores.

Use it to:

• turn scattered memory into a layered runtime
• add typed links between knowledge, entities, events, and reusable assets
• return a compact retrieval bundle for prompts, ranking, and observability
• keep write-back strict so the store stays durable instead of noisy

Quick Start

Follow this default sequence:

1. Define the two-axis memory model with layers and scopes.
2. Store knowledge_entry, knowledge_link, and entity records.
3. Build a query from role, objective, direction, and recent signals.
4. Rank candidates, expand one hop through typed links, and trim results.
5. Expose a small output bundle to prompts, task ranking, and dashboards.
6. Write back only stable findings after successful runs.

Memory Model

Use two axes.

• Layers: working, episodic, semantic, procedural, policy
• Scopes: session, shared, published

Default placement rules:

• gene, capsule, skill, and reusable playbooks belong to procedural.
• Event logs, task outcomes, and run histories belong to episodic.
• Stable conclusions and research briefs belong to semantic.
• User constraints and system rules belong to policy.

Core Records

The runtime should center on three record types:

• knowledge_entry: the main unit of stored knowledge
• knowledge_link: a typed relationship between records
• entity: the canonical form of a repo, module, topic, paper, person, org, or asset

Storage

Default files:

• memory/knowledge/knowledge_store.jsonl
• memory/knowledge/knowledge_links.jsonl
• memory/knowledge/knowledge_index.json
• memory/knowledge/entity_index.json

Retrieval Flow

When retrieval is needed:

1. Build the current query from role, objective, direction, query bundle, and signals.
2. Retrieve candidate knowledge from layered sources.
3. Expand one hop through typed links.
4. Return a compact bundle with:

- knowledge_hits - knowledge_bias_tags - linked_entities - linked_genes - memory_layers - knowledge_context_preview

Typed Links

Recommended relations:

• mentions_entity
• supports_gene
• derived_from_event
• abstracts_task
• contradicts
• supersedes
• same_topic_as
• evidence_for
• used_by_cycle

Write-Back Rule

Only write back stable, high-signal findings.

• Good: validated findings, repeated problem patterns, reusable research summaries
• Bad: raw logs, speculative notes, temporary scratch content

Adapter Surfaces

Keep the runtime decoupled from any one agent loop. Plug it into host systems through generic adapters:

• query_builder: turns role, objective, and signals into a retrieval query
• retrieval_selector: ranks hits and prepares the runtime output bundle
• task_ranker: adds knowledge relevance into task or action scoring
• prompt_context: injects a compact knowledge block into prompts
• write_back: records durable findings after successful runs
• observability: exposes hit counts, linked entities, and layer coverage to reports or dashboards

Additional Resources

Use these files:

• README.md: overview, use cases, and integration checklist
• examples.md: example retrieval, ranking, and write-back flows
• reference.md: record schemas, output shape, and adapter details