Concepts · Knowledge Graph

48 Concepts mit Relations

Der Knowledge Graph hinter den Recipes. Jedes Concept ist mit verwandten verlinkt (`REQUIRES`, `EXTENDS`, `ALTERNATIVE_TO`, `USES`, ...). Nutzbar als Navigation — oder als Lookup wenn Du wissen willst was ein Begriff genau bedeutet.

Commands · 2

mcp_installfree

CLI command that registers an MCP server in ~/.claude.json.

/rewind (Claude Code)

Built-in Claude Code command (also aliased /checkpoint) that rolls back code AND conversation to an earlier point. Session-level safety net.

Config Files · 4

Cross-tool variant of CLAUDE.md. Codex reads it. Cursor reads it. Symlink one to the other for one source of truth.

Markdown file Claude Code reads at session start as system context. The single biggest leverage point in your setup.

settings_jsonfree

~/.claude/settings.json

JSON config for hooks, permissions, env vars. User-scoped. Auto-merged with project ./.claude/settings.json.

codex_config_tomlfree

~/.codex/config.toml

Codex equivalent of Claude Code's ~/.claude.json. TOML format. MCP servers, projects, features.

Features · 7

Auto-Dream (Claude Code)

Background subagent that consolidates memory files every 24 hours. Deduplicates, removes outdated info, rewrites relative dates as absolute.

auto_memoryfree

Auto-Memory (Claude Code)

Default-on since March 2026. Claude writes notes to local memory files in the background while you work. Toggle via autoMemoryEnabled in settings.json.

Deterministic shell commands that run on Claude Code events. Hard guarantees the model cannot bypass.

memory_basicsfree

Persistent context across sessions. Default: ~/.claude/projects/<project>/memory/. Better: a memory MCP server.

permissionsfree

Allow- and deny-lists in settings.json that auto-clear or block tool calls without prompting.

Markdown files Claude loads on demand based on natural-language triggers. Like slash commands but smarter.

Markdown files in ~/.claude/agents/ that spin up an isolated agent context for specific tasks.

MCP Servers · 6

brave_search_mcpfree

brave-search MCP

Brave's own crawled search index as an MCP server. Free 2k queries/month.

codebase_memory_mcpfree

codebase-memory-mcp

Indexes your repo into a Postgres-backed knowledge graph. 14 tools.

Deep code analysis on top of a code knowledge graph. 34 tools.

Up-to-date library + framework docs as MCP tools. No API key needed.

Official GitHub MCP server. 30+ tools for issues, PRs, code search, reviews.

mcp-nex (StudioMeyer Memory)

Memory MCP server with 53 tools — sessions, learnings, decisions, knowledge graph, skills tracking.

Patterns · 24

Who did what when. Immutable record of every mutating operation.

branch_per_askfree

Every meaningful AI task gets its own Git branch. Tiny merge surface, instant rollback when bad, clean merge when good.

confidence_decayfree

Confidence decay (FadeMem)

Time-based reduction of confidence on memories not accessed recently. Keeps the graph from drowning in stale beliefs without manual archival.

context_pollutionfree

Context pollution

Memory loaded too greedily floods every prompt with marginally-relevant context. AI loses focus, misses the point, hallucinates trying to integrate everything.

dual_transportpro

Dual transport (stdio + HTTP)

One MCP server codebase that runs both as stdio subprocess (local) and Streamable HTTP (remote SaaS).

EU AI Act + GDPR

Compliance basics for small SaaS that touches EU users. Art. 52 transparency, art. 6 lawful basis.

git_safety_netfree

Git as the AI safety net

Git is mandatory when building with AI. AI changes 8 files in seconds, editor undo cannot save you, only commits give you a working rollback target.

health_endpointpro

/health endpoint

Liveness probe that monitoring + load balancers hit. Returns 200 + version + status.

idempotencyteam

Idempotency keys

Handle retries (Stripe webhooks, client retries) without double-charging or double-creating.

knowledge_graphfree

Knowledge graph

Memory stored as entities and relationships rather than flat text. Enables multi-hop reasoning ("who works at the company that supplies X").

Magic-link auth

Email-based passwordless login. User clicks a one-time URL, server exchanges it for an access token.

Model Context Protocol — open standard for connecting AI clients (Claude, Codex, Cursor) to external tools.

memory_hygienepro

Memory hygiene routine

Weekly + monthly + quarterly maintenance ritual that keeps memory useful for years instead of months. ~20 minutes per week total.

multi_tenantpro

Multi-tenant SaaS

Single deployment serves many isolated tenants. Each tenant has its own data, no leaks.

Auth flow with code-challenge + code-verifier — secure for public clients without a stored client secret.

agent_researchteam

Orchestrated research (3 agents)

Spawn Research + Critic + Analyst in parallel on a topic, synthesize 3 reports into 1 verdict.

rate_limits_pgteam

Rate limits in Postgres

Atomic upsert with case-reset. No Redis needed for low-volume SaaS.

Row Level Security (Postgres)

Database-side tenant isolation. Defense in depth on top of app-level filtering.

ai_rollback_patternfree

Safe AI rollback (no git reset --hard)

Three safe paths to undo: GUI revert, /rewind in Claude Code, or explicit checkout with named commit. Never give vague "undo it" prompts to the AI.

sm_deploy_ritualteam

SM-deploy ritual

Codified production deploy: 5 phases (orient, pre-checks parallel, go/no-go, deploy, verify). Saves you from the "I forgot to check" failure mode.

ssh_batchingteam

Avoid fail2ban self-bans on your own servers. Pattern: max 5-10 SSH connections/hour, batch with && chains.

stale_factsfree

Stale facts in memory

You change your address, role, or project. The memory keeps the old fact. Without contradiction detection, you have a hallucination source.

stripe_checkoutpro

Stripe Checkout

Hosted checkout flow for subscriptions. You redirect to Stripe, they redirect back with a session ID.

Sycophancy amplification

Documented LLM behaviour of adapting to user opinion. Memory makes it worse — the AI accumulates user preferences as facts and confirms them forever.

Services · 4

Letta (formerly MemGPT)

OS-inspired tiered memory architecture (Core + Archival + Recall). Apache 2.0, full self-host. LongMemEval not published.

Cloud Run-backed deploy platform for MCP servers. One command from repo to https URL.

Open-source memory platform with 21+ framework integrations. Apache 2.0 client, cloud SaaS proprietary. LongMemEval 49% on independent eval.

Bi-temporal knowledge graph memory backed by the open-source Graphiti framework. SOC 2 Type II + HIPAA BAA. LongMemEval 63-71% depending on setup.

Tools · 1

longmemevalfree

LongMemEval benchmark

Academic standard for measuring memory accuracy. Tests temporal queries, multi-hop reasoning, knowledge updates, casual mentions in long conversations.