Knowledge System (v1.4)

January 12, 2026 · View on GitHub

Claude Router includes a persistent knowledge system that gives you continuity across sessions and prevents context loss.

The Intelligence Stack

┌─────────────────────────────────────────┐
│  KNOWLEDGE (Memory)                      │
│  Patterns, quirks, decisions             │
│  Survives sessions & compaction          │
└──────────────────┬──────────────────────┘
                   │ informs
┌──────────────────▼──────────────────────┐
│  ROUTING (Decision-Making)               │
│  Classification, caching, model selection│
└──────────────────┬──────────────────────┘
                   │ delegates to
┌──────────────────▼──────────────────────┐
│  AGENTS (Execution)                      │
│  fast, standard, deep, orchestrator      │
└──────────────────┬──────────────────────┘
                   │ use
┌──────────────────▼──────────────────────┐
│  SKILLS (Capabilities)                   │
│  /learn, /knowledge, /route, etc.        │
└─────────────────────────────────────────┘

Each layer makes the others smarter. Knowledge isn't standalone memory - it's integrated with routing so the system acts on what it learns.

Why This Matters

The Problem: Every Claude Code session starts fresh. When you:

End a session
Hit context limits and get auto-compacted
Come back the next day

...all the understanding Claude built up about your project is gone. You end up re-explaining the same quirks, patterns, and decisions over and over.

Manual Workaround: You could manually update CLAUDE.md with project notes, but:

It's tedious to remember to do
You have to manually extract insights from conversations
Context often gets lost before you think to save it

The Knowledge System: Automatically captures and persists project understanding:

What It Captures	Example	Why It Helps
Patterns	"Error handling wraps async calls in try-catch with custom logger"	Future sessions know the codebase conventions
Quirks	"Auth service returns 200 even on errors - check response.success"	Avoids re-discovering gotchas
Decisions	"Chose TypeScript strict mode to catch bugs early"	Preserves rationale for future reference

How It Creates Continuity

With knowledge system:

Session 1: Discover auth quirk → /learn saves it
Session 2: Ask about auth → Claude already knows the quirk
Session 3: Debug auth issue → Context preserved from session 1

Without knowledge system:

Session 1: Discover auth quirk → Session ends, lost
Session 2: Re-discover auth quirk → Session ends, lost
Session 3: Re-discover auth quirk again → Frustrating loop

Knowledge Commands

Command	What It Does
`/learn`	Extract insights from current conversation NOW
`/learn-on`	Enable continuous learning (auto-extracts every 10 queries)
`/learn-off`	Disable continuous learning
`/knowledge`	View knowledge base status and recent learnings
`/learn-reset`	Clear all knowledge and start fresh

Quick Start

# After a productive session where you discovered something useful:
/learn

# Or enable continuous learning for long sessions:
/learn-on

# Check what's been learned:
/knowledge

Where It's Stored

Knowledge lives in knowledge/learnings/ within your project:

patterns.md - Approaches that work well
quirks.md - Project-specific gotchas
decisions.md - Architectural decisions with rationale

Privacy: Knowledge is gitignored by default (local only). To share with your team:

Edit knowledge/.gitignore to allow specific files
Commit the knowledge files you want to share

Advanced: Informed Routing (Opt-in)

The knowledge system can inform routing decisions. If you've learned that "auth is complex in this project," queries about auth can be routed to Opus automatically.

To enable (conservative, off by default):

# Edit knowledge/state.json and set:
# "informed_routing": true

This is conservative by design - it requires strong signals (2+ keyword matches) and uses small confidence adjustments to avoid over-routing to expensive models.

Configuration

knowledge/state.json controls behavior:

{
  "learning_mode": false,
  "informed_routing": false,
  "informed_routing_boost": 0.1,
  "extraction_threshold_queries": 10
}

Setting	Description	Default
`learning_mode`	Auto-extract insights periodically	`false`
`informed_routing`	Let knowledge influence routing	`false`
`informed_routing_boost`	Max confidence adjustment	`0.1`
`extraction_threshold_queries`	Queries between auto-extractions	`10`