docs: add Kairn memory MCP + resource evaluations + guide updates

- guide/ultimate-guide.md §10.2: Add Kairn (knowledge graph memory with biological decay) - Typed relationships (depends-on, resolves, causes), 18 MCP tools - Updated comparison table: Serena / grepai / doobidoo / Kairn - Added decision routing for long-term memory + causality tracking - guide/ultimate-guide.md §5.1: Add real-world CLAUDE.md migration example (Avo, 600-line → 15 path-scoped files) - guide/ai-ecosystem.md: Minor update - machine-readable/reference.yaml: Add Kairn entries - examples/config/mcp.json: Add Kairn MCP config - docs/resource-evaluations/: Add 2 new evaluations (context-evaluator-packmind, kairn-memory-mcp) - docs/resource-evaluations/agents-md-empirical-study: Add community reception section - docs/resource-evaluations/2026-02-23-agentsview: Minor fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-02-25 17:39:20 +01:00 · 2026-02-25 17:39:20 +01:00 · 97f9167a61
commit 97f9167a61
parent a6b0a0084a
8 changed files with 220 additions and 11 deletions
--- a/docs/resource-evaluations/2026-02-23-agentsview-session-analytics.md
+++ b/docs/resource-evaluations/2026-02-23-agentsview-session-analytics.md
@ -143,4 +143,4 @@ The challenge agent recommended **3/5** (vs initial 4/5) for the following reaso
 - **Final score**: **3/5**
 - **Action**: **Integrate** — `observability.md` + mention in `third-party-tools.md`
 - **Timing**: 2-4 weeks (wait for repo to reach ~200+ stars)
- **Confidence**: **Medium** — real gap confirmed, tool verified functional, but adoption too recent for high confidence. Wes McKinney's credibility (pandas) is a strong positive signal.
+- **Confidence**: **Medium** — real gap confirmed, tool verified functional, but adoption too recent for high confidence. Wes McKinney's credibility (pandas) is a strong positive signal. 
--- a/docs/resource-evaluations/2026-02-25-context-evaluator-packmind.md
+++ b/docs/resource-evaluations/2026-02-25-context-evaluator-packmind.md
@ -0,0 +1,85 @@
+# Evaluation: Context-Evaluator (context-evaluator.ai)
+
+**Date**: 2026-02-25
+**Type**: URL + texte (LinkedIn post de Cédric Teyton, CTO Packmind)
+**Source**: https://context-evaluator.ai/
+
+---
+
+## Resume du contenu
+
+- **Outil open-source** (Apache-2.0) par Packmind (même équipe que coding-agents-matrix déjà dans le guide)
+- **Scanner de fichiers CLAUDE.md, AGENTS.md, copilot-instructions.md** avec 17 évaluateurs spécialisés (13 détecteurs d'erreurs + 4 générateurs de suggestions)
+- **Multi-agent**: Claude Code, Cursor, GitHub Copilot, OpenCode, OpenAI Codex
+- **Dual interface**: CLI local + web UI (context-evaluator.ai)
+- **Stack**: Bun, React, Tailwind CSS, TypeScript
+- **GitHub**: PackmindHub/context-evaluator | 8 stars | 2 contributeurs | v0.3.0 (23 fev 2026)
+
+---
+
+## Score de pertinence: 3/5
+
+**Justification**: Premier outil qui formalise des critères de qualité pour les fichiers CLAUDE.md avec des checks reproductibles. Valeur pédagogique des 17 critères même indépendamment de l'outil. Mais v0.3.0 avec 8 stars = side project expérimental, risque d'abandon réel.
+
+---
+
+## Comparatif
+
+| Aspect | Context-Evaluator | Notre guide |
+|--------|------------------|-------------|
+| Audit CLAUDE.md | 17 évaluateurs automatisés (déterministe) | audit-prompt.md (LLM-dependent, subjectif) |
+| Scope | Multi-agent (5 outils) | Claude Code spécifique |
+| Maintenance context files | Détection erreurs + suggestions | Documentation des bonnes pratiques (section 3.1) |
+| Maturité | v0.3.0, 8 stars, expérimental | 20K lignes, 175 templates, établi |
+| Remediation | Auto-fix via AI intégré | Templates + checklists manuelles |
+| Context drift detection | Oui (mismatch code/docs) | Mentionné en concept (iterative-refinement.md:253) |
+
+---
+
+## Recommandations
+
+**Action**: Intégrer comme mention légère (pas de section dédiée)
+
+1. **`guide/ai-ecosystem.md` ligne ~2064** (section Packmind Related Resources): Ajouter context-evaluator à côté de coding-agents-matrix. 3-5 lignes max.
+2. **`machine-readable/reference.yaml`**: Ajouter entrée `context_evaluator` pointant vers ai-ecosystem.md
+3. **Ne PAS créer de section dédiée** à 8 stars et v0.3.0
+4. **Optionnel**: Lien depuis section maintenance CLAUDE.md vers ai-ecosystem pour les curieux
+
+**Priorité**: Basse (intégrer quand opportun)
+
+---
+
+## Challenge (technical-writer)
+
+- **Score ajusté**: 3/5 confirmé, justification retravaillée
+- **Points manqués**: Les 17 critères concrets (le vrai trésor), le modèle de scoring, le dogfooding potentiel sur notre propre CLAUDE.md
+- **Risques de non-intégration**: Quasi nuls. 8 stars = personne ne reprochera l'absence. Le risque inverse (contenu mort si outil abandonné) est plus réel.
+- **Vraie valeur ajoutée**: Les critères formels d'évaluation, pas l'outil lui-même
+- **Suggestion pertinente**: Faire tourner l'outil sur notre propre CLAUDE.md pour data factuelle
+
+---
+
+## Fact-Check
+
+| Affirmation | Vérifiée | Source |
+|-------------|----------|--------|
+| Open-source Apache-2.0 | OK | GitHub repo confirmé |
+| Par Packmind (Cédric Teyton) | OK | GitHub + ai-ecosystem.md:2012 déjà référencé |
+| 17 évaluateurs (13+4) | OK | Page web + README GitHub |
+| Supporte Claude Code, Cursor, Copilot, OpenCode, Codex | OK | Page web |
+| v0.3.0, release 23 fev 2026 | OK | GitHub releases |
+| 8 stars, 2 contributeurs | OK | GitHub repo (au 25 fev 2026) |
+| Gratuit ($0) | OK | Page web |
+| "An AI coding agent is only as smart as the last time your context was reviewed" | Claim marketing, non vérifiable | N/A |
+
+**Perplexity**: Aucun résultat spécifique trouvé sur context-evaluator.ai. Outil trop récent/niche pour avoir de la couverture presse.
+
+**Corrections**: Aucune hallucination détectée. Le texte LinkedIn est fidèle aux faits du repo.
+
+---
+
+## Decision finale
+
+- **Score final**: 3/5
+- **Action**: Intégrer comme mention dans ai-ecosystem.md section Packmind + entrée reference.yaml
+- **Confiance**: Haute (facts vérifiés, repo existe, même auteur que ressource déjà intégrée)
--- a/docs/resource-evaluations/2026-02-25-kairn-memory-mcp.md
+++ b/docs/resource-evaluations/2026-02-25-kairn-memory-mcp.md
@ -0,0 +1,47 @@
+# Evaluation: Kairn — Knowledge Graph Memory MCP
+
+**Date**: 2026-02-25
+**Evaluator**: Claude (eval-resource skill)
+**Score**: 4/5 — Intégré
+
+---
+
+## Sources évaluées
+
+1. **LinkedIn post** — Robin Lorenz, "Context Engineering for Claude Code", 24 fév. 2026
+   - Score: 2/5 — Watch list (claims invérifiables, pas de code source)
+2. **GitHub repo** — [kairn-ai/kairn](https://github.com/kairn-ai/kairn) (MIT, Python 100%)
+   - Score: 4/5 — **Intégré**
+
+---
+
+## Résumé Kairn
+
+MCP server Python offrant une mémoire organisée en knowledge graph avec 18 outils.
+
+**Différenciateurs clés vs Serena/doobidoo**:
+- Typed relationships (`depends-on`, `resolves`, `causes`)
+- Biological decay model : solutions ~200j, workarounds ~50j (auto-pruning)
+- Full-text search + confidence routing
+- Cross-session et cross-IDE
+
+---
+
+## Décision d'intégration
+
+| Élément modifié | Description |
+|-----------------|-------------|
+| `guide/ultimate-guide.md` ~10233 | Nouvelle section "Kairn: Knowledge Graph Memory with Biological Decay" |
+| `guide/ultimate-guide.md` | Comparison Matrix + "When to use" table — colonne Kairn ajoutée |
+| `guide/ultimate-guide.md` | Doobidoo limitation "No expiration" → pointe vers Kairn |
+| `examples/config/mcp.json` | Entrée `kairn` ajoutée |
+
+---
+
+## Watch list — LinkedIn Lorenz
+
+**Score: 2/5** — Réévaluer si un repo/article technique accompagne le post.
+
+Raison du rejet : post LinkedIn sans code source, stats non vérifiables (93% réduction tokens), Perplexity n'a trouvé aucune trace indépendante. La philosophie "infrastructure outlives prompts" est déjà couverte dans `guide/methodologies.md`.
+
+**Réévaluation trigger** : publication d'un repo GitHub accompagnant le post.
--- a/docs/resource-evaluations/agents-md-empirical-study-2602-11988.md
+++ b/docs/resource-evaluations/agents-md-empirical-study-2602-11988.md
@ -127,4 +127,16 @@ Ajouter une mise en garde :

 ---

+---
+
+## 📣 Réception communautaire
+
+La communauté a simplifié les résultats en "delete your CLAUDE.md" (cf. [Charly Wargnier, LinkedIn](https://www.linkedin.com/posts/charlywargnier_everyone-is-screaming-delete-your-claudemd-activity-7431988275193622528-pfBW)).
+
+**Nuance importante à intégrer dans le guide** : la commande `/init` génère un context file LLM-generated → c'est ce type de fichier qui dégrade les performances (-3%). Les fichiers écrits manuellement restent bénéfiques (+4%).
+
+D'autres posts (ex: [Daniel Vikulov, LinkedIn](https://www.linkedin.com/)) paraphrasent fidèlement l'étude sans valeur ajoutée — ne pas les citer comme sources indépendantes.
+
+---
+
 *Évaluation effectuée le 2026-02-19 | Méthode: WebFetch + Perplexity + grepai_search + agent challenge*
--- a/examples/config/mcp.json
+++ b/examples/config/mcp.json
@ -38,6 +38,11 @@
      "command": "memory",
      "args": ["server"],
      "description": "Semantic memory with cross-session search (complements Serena)"
+    },
+    "kairn": {
+      "command": "python",
+      "args": ["-m", "kairn", "serve"],
+      "description": "Knowledge graph memory with biological decay — auto-expires stale info (complements doobidoo/Serena)"
    }
  }
 }
--- a/guide/ai-ecosystem.md
+++ b/guide/ai-ecosystem.md
@ -2063,6 +2063,7 @@ An **interactive comparison matrix** of 23 AI coding agents across 11 technical

 - [Packmind](https://packmind.com): Context engineering & governance for AI coding agents
 - [Packmind OSS](https://github.com/PackmindHub/packmind): Framework for versioning AI coding context
+- [Context-Evaluator](https://context-evaluator.ai) ([GitHub](https://github.com/PackmindHub/context-evaluator)): Open-source scanner for CLAUDE.md / AGENTS.md / copilot-instructions.md — 17 evaluators (13 error detectors + 4 suggestion generators), supports Claude Code, Cursor, Copilot, OpenCode, Codex. CLI + web UI. Apache-2.0, v0.3.0 (Feb 2026).
 - [Claude Code Templates](https://github.com/davila7/claude-code-templates): 200+ templates for Claude Code (17k⭐)
 - [Awesome Claude Code](https://github.com/hesreallyhim/awesome-claude-code): Curated tool library

--- a/guide/ultimate-guide.md
+++ b/guide/ultimate-guide.md
@ -5171,7 +5171,7 @@ These rules only apply when working with API files:
 - Include rate limiting middleware
 ```

-This enables progressive context loading—rules only appear when Claude works with matching files.
+This enables progressive context loading—rules only appear when Claude works with matching files. Real-world example: Avo migrated a 600-line CLAUDE.md to ~15 path-scoped files, reporting sharper responses and easier maintenance across domains. ([Björn Jóhannsson](https://www.linkedin.com/posts/bj%C3%B6rn-j%C3%B3hannsson-72435083_your-claudemd-is-eating-your-context-window-activity-7431750526729338881-ODSs))

 **How matching works**:
 - Patterns use glob syntax (same as `.gitignore`)
@ -10231,6 +10231,50 @@ Device C ──┘

 > **Source**: [doobidoo/mcp-memory-service GitHub](https://github.com/doobidoo/mcp-memory-service) (791 stars, v10.0.2)

+### Kairn: Knowledge Graph Memory with Biological Decay
+
+> **⚠️ Status: Under Testing** - Evaluated Feb 2026. MIT licensed, Python 100%. Feedback welcome!
+
+**Purpose**: Long-term project memory organized as a knowledge graph with automatic decay — stale information expires on its own, preventing context pollution.
+
+**Key differentiators vs doobidoo/Serena**:
+- **Typed relationships**: `depends-on`, `resolves`, `causes` — captures causality, not just content
+- **Biological decay model**: solutions persist ~200 days, workarounds ~50 days — auto-pruning without `delete_memory` calls
+- **18 MCP tools**: graph ops, project tracking, experience management, intelligence layer (full-text search, confidence routing, cross-workspace patterns)
+
+| Feature | Serena | doobidoo | Kairn |
+|---------|--------|----------|-------|
+| Storage model | Key-value | Semantic embeddings | Knowledge graph |
+| Memory decay / auto-expiry | No | No | Yes (biological) |
+| Typed relationships | No | Tags only | depends-on / resolves / causes |
+| Full-text search | No | Yes | Yes |
+| Auto-pruning stale info | No | No | Yes |
+
+**When Kairn makes sense**:
+- Long-running projects where workarounds from months ago become noise
+- When causality matters: "this breaks *because* of that", "this fix *resolves* that bug"
+- Teams wanting automatic knowledge hygiene without manual cleanup
+
+**MCP Config**:
+
+```json
+"kairn": {
+  "command": "python",
+  "args": ["-m", "kairn", "serve"],
+  "description": "Knowledge graph memory with biological decay"
+}
+```
+
+**Install**:
+
+```bash
+pip install kairn
+# or from source:
+git clone https://github.com/kairn-ai/kairn && cd kairn && pip install -e .
+```
+
+> **Source**: [kairn-ai/kairn GitHub](https://github.com/kairn-ai/kairn) (MIT, Python 100%)
+
 ### MCP Memory Stack: Complementarity Patterns

 > **⚠️ Experimental** - These patterns combine multiple MCP servers. Test in your workflow before relying on them.
@ -10257,14 +10301,15 @@ Device C ──┘

 **Comparison Matrix**:

-| Capability | Serena | grepai | doobidoo |
-|------------|--------|--------|----------|
-| Cross-session memory | Key-value | No | Semantic |
-| Cross-IDE memory | No | No | Yes |
-| Cross-device sync | No | No | Yes (Cloudflare) |
-| Knowledge Graph | No | Call graph | Decision graph |
-| Fuzzy search | No | Code | Memory |
-| Tags/categories | No | No | Yes |
+| Capability | Serena | grepai | doobidoo | Kairn |
+|------------|--------|--------|----------|-------|
+| Cross-session memory | Key-value | No | Semantic | Knowledge graph |
+| Cross-IDE memory | No | No | Yes | Yes |
+| Cross-device sync | No | No | Yes (Cloudflare) | No |
+| Knowledge Graph | No | Call graph | Decision graph | Typed relationships |
+| Fuzzy search | No | Code | Memory | Full-text + semantic |
+| Tags/categories | No | No | Yes | Yes |
+| Memory decay / auto-expiry | No | No | No | Yes (biological) |

 **Usage Patterns**:

@ -10336,6 +10381,8 @@ retrieve_memory("work in progress?")
 | "Share across devices" | doobidoo + Cloudflare | Cloud sync |
 | "Code symbol location" | Serena `find_symbol()` | Code indexation |
 | "Code by intent" | grepai `search()` | Semantic code search |
+| "Long-term project memory, auto-expiry" | Kairn | Biological decay model |
+| "Why did X break / what resolved Y?" | Kairn | Typed relationships (resolves, causes) |

 **Current Limitations** (doobidoo):

@ -10344,7 +10391,7 @@ retrieve_memory("work in progress?")
 | No versioning | Can't see decision history | Include dates in content |
 | No permissions | Anyone can modify | Use separate DBs per team |
 | No source linking | No link to file/line | Include file refs in content |
-| No expiration | Stale memories persist | Manual cleanup with `delete_memory` |
+| No expiration | Stale memories persist | Manual cleanup with `delete_memory` OR use Kairn (auto-decay) |
 | No git integration | No branch-aware memory | Tag with branch name |

 </details>
--- a/machine-readable/reference.yaml
+++ b/machine-readable/reference.yaml
@ -1269,6 +1269,18 @@ ecosystem:
      agents_snapshot: 23  # As of 2026-01-19 (evolves over time)
      positioning: "Discovery tool - Use Matrix to find/compare agents, use this guide to master Claude Code"
      note: "External resource - verify data freshness as agents/criteria evolve"
+    context_evaluator:
+      url: "context-evaluator.ai"
+      github: "github.com/PackmindHub/context-evaluator"
+      maintainer: "Packmind (Cédric Teyton)"
+      license: "Apache-2.0"
+      focus: "Scanner for CLAUDE.md/AGENTS.md/copilot-instructions.md — 17 evaluators (13 error detectors + 4 suggestion generators)"
+      updated: "2026-02-23"
+      tech_stack: "Bun, React, Tailwind CSS, TypeScript"
+      supports: ["Claude Code", "Cursor", "GitHub Copilot", "OpenCode", "OpenAI Codex"]
+      interface: ["CLI", "Web UI"]
+      guide_section: "ai-ecosystem.md#related-resources"
+      note: "v0.3.0, experimental (8 stars as of 2026-02-25) — mention only, no dedicated section"
  practitioner_insights:
    dave_van_veen:
      url: "davevanveen.com/blog/agentic_coding/"