Research context: validating the "persistent knowledge docs" hypothesis from the Rhapsody experiment. The question is whether a persistent, agent-readable knowledge artifact can serve four use cases: backfill, active authoring, self-healing, and team process encoding.
- Backfill — bootstrapping docs from existing code/history (what's already true but undocumented)
- Active authoring — agents writing knowledge as they work (decisions, conventions, architecture)
- Self-healing — detecting and correcting drift when docs become stale
- Team processes — encoding "how we work" (reviewers, channels, rituals) as agent-readable knowledge
Anthropic — Effective Context Engineering for AI Agents https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
Defines context engineering as curating what goes into the context window each inference. Frames the shift from "write a good prompt" to "build a system that assembles the right context dynamically."
Simon Willison — Context Engineering https://simonwillison.net/2025/Jun/27/context-engineering/
Shorter take on the same shift. Compares it to deciding which sticky notes to leave on a junior dev's desk — and when to clear the desk so only actionable context remains.
Google Cloud — Open Knowledge Format (OKF) v0.1 Spec https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf
Vendor-neutral, open spec (June 2026). A knowledge bundle is a directory of Markdown files. Each file encodes one concept, with YAML frontmatter (type is the only required field). Apache 2.0. Minimal contract — deliberately doesn't enforce tooling or schema registries.
Relevant because: this is what "knowledge docs as an open format" looks like when someone standardizes it.
Codified Context: Infrastructure for AI Agents in a Complex Codebase https://arxiv.org/html/2602.20478v1
Academic paper arguing that single-file manifests (AGENTS.md) don't scale. Proposes tiered infrastructure: hot memory (conventions, hooks), domain-expert agents, cold memory (knowledge base with dozens of persistent docs). Introduces freshness/decay signals so agents know when to re-validate.
Relevant because: directly addresses the scaling problem and the self-healing question.
Anthropic — Claude Agent "Dreaming" https://www.aitrove.ai/blog/code-with-claude-2026-dreams-routines-agents
Agents analyze prior runs between sessions, revise their own memory files (MEMORY.md), prune stale knowledge, and add new learnings. Can be auto-approved or human-supervised. This is Anthropic's answer to "how do docs stay fresh."
Anthropic — Context Management on Claude Platform https://claude.com/blog/context-management
Automatic context pruning: stale tool calls/results are removed as sessions approach token limits. Complements the dreaming approach (dreaming = between-run maintenance, pruning = within-run maintenance).
Claude Code & Agent Memory: Best Practices 2026 https://orchestrator.dev/blog/2026-04-06--claude-code-agent-memory-2026/
Documents the four-layer memory architecture:
- CLAUDE.md — static, human-authored truths
- Auto Memory (MEMORY.md) — dynamic, learned from corrections
- Memory Tool directory — rich persistent files
- Subagent Memory — per-subagent isolated stores
Relevant because: shows how one system separates "what humans assert" from "what the agent learns" from "what persists long-term."
Swimm — Integrating Continuous Documentation https://swimm.io/blog/integrating-swimm-and-continuous-documentation-into-your-workflow
Code-coupled docs that reference live code snippets. If code changes, Swimm auto-syncs or flags the doc for review. "Docuthon" model for bootstrapping docs in undocumented codebases.
Relevant because: their auto-sync (code hash → doc validity) is a concrete self-healing mechanism. The "docuthon" is a backfill strategy.
Karpathy's LLM Wiki Guide https://blog.starmorph.com/blog/karpathy-llm-wiki-knowledge-base-guide
How to build AI-maintained knowledge bases that agents ingest and update, replacing RAG with plain Markdown. Argues for canonical Markdown as the source of truth, with embeddings/graphs as derived views.
GitHub — Building an Agentic Memory System for Copilot https://github.blog/ai-and-ml/github-copilot/building-an-agentic-memory-system-for-github-copilot/
How Copilot Memory works: user/repo/session scopes, cross-agent sharing (code review agent reads memories from coding agent), opt-in persistence.
Long-term Memory for AI Agents: Mem0, LangGraph, and Beyond https://thread-transfer.com/blog/2025-03-23-long-term-memory-ai-agents/
Covers procedural memory (how-to-do-things) as distinct from episodic (what happened) and semantic (facts). Relevant framing for team processes — rituals and workflows are procedural memory.
Anthropic — Advanced Tool Use https://www.anthropic.com/engineering/advanced-tool-use
Programmatic tool calling in sandboxed environments that don't fill the model's context. Relevant to keeping knowledge doc interactions lightweight.
| Front | Industry approach | Gap we'd fill |
|---|---|---|
| Backfill | Swimm docuthons, LLM-generated summaries, code-coupled refs | Agent that reads git history + existing docs and bootstraps knowledge per-concept |
| Active authoring | OKF (one concept per file), CLAUDE.md/MEMORY.md split | Agent writes knowledge as a side-effect of doing work (Scribe model) |
| Self-healing | Anthropic "dreaming," Swimm auto-sync, code-hash freshness | Between-run review loop that detects drift and proposes updates |
| Team processes | .cursorrules, copilot-instructions.md, OKF frontmatter | Human-AI-readable "how this team works" artifact that agents can propose updates to |
The common thread: everyone is converging on Markdown + YAML frontmatter as the format, one concept per file as the granularity, and separation of human-authored vs agent-learned as the key architectural decision.
What nobody has cracked yet: a closed loop where the agent both uses and maintains the knowledge without human approval becoming a bottleneck. Anthropic's "dreaming" is closest but still requires periodic human review.