/swafra — Knowledge Graph Memory

swafra gives Claude persistent, searchable memory using a local knowledge graph. No cloud, no GPU, no API keys.

GitHub: https://github.com/kunal12203/swafra Benchmark: 94.7% recall_all@10 on LongMemEval-S (beats Supermemory 95% end-to-end QA on retrieval)

---

Setup (one time)

git clone https://github.com/kunal12203/swafra
cd swafra
pnpm install && pnpm build

pip install fastembed numpy  # optional — falls back to local embedder without it

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

{
  "mcpServers": {
    "swafra": {
      "command": "node",
      "args": ["/absolute/path/to/swafra/dist/index.js"]
    }
  }
}

Add to Claude Code (.claude/settings.json):

{
  "mcpServers": {
    "swafra": {
      "command": "node",
      "args": ["/absolute/path/to/swafra/dist/index.js"]
    }
  }
}

Restart Claude after adding the config.

---

Tools

Tool	When to use
add_knowledge	Store a document, note, or any text into memory
search_knowledge	Find relevant chunks by natural language query
get_context	Best for retrieval — search + graph walk combined
graph_walk	Explore context around a specific chunk
list_sources	See what's been stored
delete_source	Remove a source from memory

---

What You Must Do When Invoked

When the user types /swafra, read their request and do ONE of:

Store something

If the user wants to remember/store something:

add_knowledge(text="<the content>", title="<short label>")

Confirm: "Stored as '' — searchable now."

Retrieve something

If the user asks a question or wants to find something:

get_context(query="<their question>", k=10)

Use the returned chunks to answer. If nothing relevant comes back, say so clearly — don't hallucinate.

List what's stored

list_sources()

Show the list cleanly.

Delete something

delete_source(title="<title>")

Confirm deletion.

---

Usage examples

/swafra store this meeting transcript: [paste text]
/swafra what did we decide about the API design?
/swafra what do you know about the auth system?
/swafra list everything stored
/swafra forget the meeting from last Tuesday

---

How it works

1. Leiden chunking — splits text into semantically coherent chunks via community detection on a hybrid graph (semantic similarity + entity co-occurrence + position)
2. Local embeddings — fastembed (ONNX, CPU) or deterministic hash vectors — no API calls
3. Fused retrieval — BM25 + vectors + entity overlap + character n-grams, source-diverse ranking
4. Knowledge graph — chunks are connected with sequential, similarity, and community edges for graph-walk retrieval

Storage: ~/.swafra/ JSON files. No database required.