KNDL — Knowledge Node Data Link

The format Anthropic Memory was waiting for

→ kndl.artdaw.com — live docs, protocol reference, interactive examples, explorer

→ watch about KNDL on YouTube

---

Anthropic just shipped Memory for agents — a filesystem. But filesystems are dumb about confidence, time, and source. Agents fill them with markdown that can't be queried, won't decay, and loses provenance the moment it's written.

KNDL is the format that makes Memory actually smart.

Anthropic Memory  =  WHERE   filesystem, persistence, permissions
KNDL              =  WHAT    the format of files Claude writes
kndl-mcp / CLI    =  HOW     query, decay, contradiction detection, provenance

---

Get started in 60 seconds

git clone https://github.com/artdaw/kndl
cd kndl/packages/kndl-memory
pnpm install && pnpm build

Add to Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "kndl": {
      "command": "node",
      "args": ["/path/to/kndl/packages/kndl-memory/dist/server.js"],
      "env": { "KNDL_STORAGE": "sqlite:///Users/you/.kndl/memory.db" }
    }
  }
}

If you use nvm, specify the full node path so Claude Desktop uses the right version: "command": "/Users/you/.nvm/versions/node/v24.9.0/bin/node"

Restart Claude Desktop. Now ask: "Remember that Alice is a staff engineer with confidence 0.95."

---

Why not just markdown?

	Markdown	KNDL
Know when a fact went stale	✗	✅ decay + effective_confidence
Surface contradictions between sources	✗	✅ ranked by recency + confidence
Trace a claim to its origin	✗	✅ source URI + derivedFrom chain
Time-travel ("what did we believe last week?")	✗	✅ as_of bitemporal query
Open-world negation ("no known allergy")	✗	✅ negated: true
PHI / PII sensitivity gating	✗	✅ classification filtered by default

---

The fact shape

One immutable file per assertion. Every fact is a JSON-LD document:

{
  "@context": "https://kndl.artdaw.com/context/v1.jsonld",
  "@id":       "fact:alice-role-20260426t100000z-ab12cd34",
  "@type":     "Fact",
  "statement":  "Alice is a staff engineer on the payments team",
  "subject":    "person:alice",
  "predicate":  "role",
  "object":     "staff engineer, payments",
  "confidence": 0.95,
  "decay":      "0.5/180d",
  "source":     "human://gleb",
  "validFrom":  "2026-04-26T10:00:00Z",
  "recordedAt": "2026-04-26T10:00:00Z"
}

decay: "0.5/180d" — confidence halves every 180 days. Facts are immutable. Updates use supersedes; history preserved for time-travel.

---

MCP tools (23)

Fact memory (11)

Tool	What it does
assert_fact	Write a new immutable fact
query_facts	Read active facts with decay-adjusted confidence. Use subject for exact match or text for substring search.
contradictions	Find conflicting facts, ranked by recency + confidence
supersede_fact	Replace a fact — history preserved for time-travel
as_of	Bitemporal: what did memory believe at timestamp X?
provenance_chain	Walk derivedFrom + supersedes backward
subscribe / unsubscribe / list_subscriptions	Fact change notifications
sync_memory_store	Sync a configured remote — direction: "pull" | "push" | "both"
list_memory_stores	List configured remotes + last-sync timestamps

Memory Store management (12, all require ANTHROPIC_API_KEY)

Tool	What it does
create_memory_store	Create a new Anthropic Memory Store
list_all_stores	List all Memory Stores in your account
get_memory_store	Get details of a store by ID
update_memory_store	Rename or update a store
delete_memory_store	Permanently delete a store
archive_memory_store	Archive a store (soft delete)
list_memories	List memories in a store (supports path_prefix filter)
get_memory	Get a memory by ID (returns full content)
create_memory	Create a memory at a given path
update_memory	Update memory content or path
delete_memory	Delete a memory

---

Storage

KNDL_STORAGE	Backend	Use case
fs:/memory	Filesystem .fact.json files	Anthropic Memory mount
sqlite:./kndl.db	SQLite WAL	Claude Desktop standalone ← default
duckdb:./kndl.duckdb	DuckDB columnar	Analytical workloads
supabase:<url>?key=<anon>	Supabase + RLS	Multi-tenant cloud

---

CLI

After pnpm build, the binary is at dist/cli.js. Make it available as kndl:

# Option A — link globally
cd packages/kndl-memory && npm link

# Option B — shell alias
alias kndl="node /path/to/kndl/packages/kndl-memory/dist/cli.js"

# Option C — run directly
node packages/kndl-memory/dist/cli.js help

Fact commands

export KNDL_STORAGE=sqlite:./kndl.db

kndl add \
  --statement "Alice is a staff engineer, payments" \
  --subject person:alice --predicate role \
  --confidence 0.95 --source "human://gleb" \
  --decay "0.5/180d" --valid-from now

kndl query --subject person:alice        # exact subject match
kndl query --text alice                  # substring search
kndl contradict --subject person:alice
kndl as-of 2026-01-01T00:00:00Z --subject person:alice
kndl provenance --id fact:alice-role-...

Remote sync commands

# Pull-only remote
kndl remote add --provider anthropic --store-id store_abc --label personal
kndl remote pull personal

# Push-enabled remote
kndl remote add --provider anthropic --store-id store_abc --label work --push
kndl add ... --tags push-to-anthropic   # tag facts to push
kndl remote push work
kndl remote sync work   # pull + push
kndl remote ls

Push selects facts tagged push-to-anthropic and skips classified data by default. Memories are stored at /kndl-facts/{fact-id} in the Memory Store.

Memory Store management (requires ANTHROPIC_API_KEY)

# Store CRUD
kndl store create  --name "My Store" --description "optional"
kndl store ls      [--archived]
kndl store get     <store_id>
kndl store update  <store_id> --name "New Name"
kndl store delete  <store_id>
kndl store archive <store_id>

# Memory CRUD (filesystem-style paths)
kndl memory create  --store <store_id> --path /notes/alice.md --content "Alice is a staff engineer"
kndl memory ls      --store <store_id> [--prefix /notes/]
kndl memory get     --store <store_id> <memory_id>
kndl memory update  --store <store_id> <memory_id> --content "Updated content"
kndl memory delete  --store <store_id> <memory_id>

---

Use with Anthropic Memory (Skill)

cp -r skills/kndl-memory/SKILL.md   ./memory/skills/
cp -r skills/kndl-memory/context/   ./memory/context/

---

HTTP server (multi-agent)

KNDL_STORAGE=sqlite:./shared.db node packages/kndl-memory/dist/server.js --http
# → http://localhost:8000/mcp

LOG_LEVEL=DEBUG ... node dist/server.js --http   # per-request debug logging

---

Repository

packages/kndl-memory/     @kndl/memory npm package (TypeScript, Node >=22)
  src/
    core.ts               decay math, fact construction, query algorithms
    types.ts              Fact, FactInput, QueryOptions, FactStore interface
    stores/               fs · sqlite · duckdb · supabase + makeStore()
    remote/
      types.ts            MemoryStore, Memory, MemoryVersion, MemoryStoreClient
      anthropic.ts        AnthropicMemoryClient (full API) + FakeMemoryStoreClient
      sync.ts             pull() · push() · syncBoth()
      config.ts           ~/.kndl/remotes.json management
    server.ts             kndl-memory-mcp (stdio + HTTP, 23 tools)
    cli.ts                kndl CLI
  tests/                  43 passing tests

skills/kndl-memory/       Claude Skill bundle
  SKILL.md                drop into /memory/skills/
  context/v1.jsonld        JSON-LD @context
  examples/               8 domain bundles · 42 facts

website/                  kndl.artdaw.com (React + Vite, 7 pages)

---

Eval

export ANTHROPIC_API_KEY=sk-ant-...
make publish-eval

KNDL must beat vanilla on ≥ 70% of 33 questions to ship.

---

License

MIT — kndl.artdaw.com