Updated 17 October 2026 · Knowledge systems and documentation
Abstract. AI-native documentation is not automated prose production. It is the engineering of organizational memory: evidence is captured, linked, versioned, retrieved, challenged and converted into reusable playbooks. Tools such as Obsidian can provide a human-readable knowledge surface; RAG and graphs provide machine retrieval; agents provide maintenance capacity. Governance determines whether the result is knowledge or merely accumulated text.
From documents to a knowledge operating system
A durable knowledge system has four layers: immutable sources, curated claims, relational structure and operational playbooks. A meeting transcript is a source; “service X has a 30-minute recovery objective” is a claim with owner and validity; its dependencies form a graph; the incident procedure is a playbook. Conflating these layers allows generated summaries to overwrite primary evidence.
MemGPT framed memory management as an operating-system problem for language models (Packer et al., 2023). RAPTOR showed that hierarchical summaries can support retrieval across levels of abstraction (Sarthi et al., 2024). Together they suggest that “more context” is not the objective; controlled movement between memory tiers is.
Obsidian as a human-governed knowledge layer
Obsidian’s local Markdown files, links and graph view make it useful as an editorial surface because the canonical artifacts remain portable (Obsidian, 2026). The graph is not automatically a knowledge graph: links need semantics, claims need provenance and templates need governance. A practical note schema includes type, status, owner, source, created date, reviewed date, validity window, related entities and sensitivity.
One durable idea, decision or claim with source and scope.
Primary material, immutable identifier and extraction context.
Context, options, decision, consequences and review trigger.
Trigger, prerequisites, steps, checks, escalation, rollback and owner.
Agentic documentation loops
- Capture: ingest approved meetings, issues, code changes and incidents with source identity.
- Extract: propose entities, claims, decisions and unresolved questions.
- Verify: link every material claim to evidence; mark conflict and uncertainty.
- Curate: a human owner approves durable knowledge and sensitivity.
- Publish: build audience-specific views without changing canonical facts.
- Revalidate: schedule review by expiry, source change or operational contradiction.
Agent S uses experience-augmented hierarchical planning, retrieving both external knowledge and internal experience (Agashe et al., 2024). The design opportunity is powerful, but experience must not become unreviewed folklore. Write back only validated lessons with scope and counterexamples.
Living playbooks for Codex, Claude, Hermes and OpenClaw
A living playbook is both readable guidance and an executable contract. Codex or Claude can follow repository-scoped instructions; Hermes and OpenClaw can apply persistent skills across surfaces. The same governance applies: version the playbook, test it against scenarios, restrict its tools, record each execution and require approval for consequential actions. The agent should report deviations rather than silently improvise.
Developer-provided context is becoming a first-class artifact. A 2025 empirical study found repeated categories of conventions, guidelines, project information, directives and examples across agent rule files (Jiang & Nam, 2025). This supports a layered documentation strategy: global principles, repository rules, directory-local constraints and task-specific briefs.
Scientific storytelling
Storytelling remains essential because evidence without structure is difficult to act on. Scientific storytelling separates observation, inference and recommendation; reports uncertainty; and makes counterevidence visible. An AI may draft the narrative, but the knowledge system should be able to answer: Which source supports this sentence? Was the source current? What alternative interpretation was considered? Who accepted the conclusion?
Quality model
| Dimension | Question | Metric |
|---|---|---|
| Provenance | Can each material claim be traced? | Supported-claim ratio |
| Freshness | Is knowledge valid for the decision date? | Expired/ownerless claim rate |
| Retrievability | Can the right audience find it? | Task-based retrieval success |
| Consistency | Are contradictions exposed? | Unresolved conflict age |
| Actionability | Does guidance produce correct outcomes? | Playbook execution success |
Knowledge principle. Do not ask an agent to “remember everything.” Engineer what deserves to become memory, preserve why it was believed and design how it can be corrected.
References
- Agashe, S., Han, J., Gan, S., Yang, J., Li, A., & Wang, X. E. (2024). Agent S: An open agentic framework that uses computers like a human. arXiv. https://arxiv.org/abs/2410.08164
- Jiang, S., & Nam, D. (2025). An empirical study of developer-provided context for AI coding assistants in open-source projects. arXiv. https://arxiv.org/abs/2512.18925
- National Institute of Standards and Technology. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). https://doi.org/10.6028/NIST.AI.600-1
- Obsidian. (2026). Obsidian help. https://help.obsidian.md/
- Packer, C., Fang, V., Patil, S. G., Lin, K., Wooders, S., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as operating systems. arXiv. https://arxiv.org/abs/2310.08560
- Sarthi, P., Abdullah, S., Tuli, A., Khanna, S., Goldie, A., & Manning, C. D. (2024). RAPTOR: Recursive abstractive processing for tree-organized retrieval. arXiv. https://arxiv.org/abs/2401.18059