Help with Persistent Memory

[Cris9870]

After using the run command with the memory name, when checking memory stats it does not appear.

The flow I am using is quite simple:

1. I install Claude and Claude-flow globally via npm.

2. Then I launch the project
   $claude-flow hive-mind spawn "Develop a to-do list application"
   --agents 10
   --strategy parallel
   --memory-namespace todolist
   --claude

3. Checking $claude-flow memory stats I get:
   Memory Bank Statistics:
   Total Entries: 0
   Namespaces: 0
   Size: 0.00 KB

4. Similarly $ claude-flow memory list
   ⚠️ No namespaces found

5. I only get a result if I explicitly create the namespace with a command like:
   $ claude-flow memory store "project-context" "Full-stack app requirements"
   ✅ Stored successfully
   📝 Key: project-context
   📦 Namespace: default
   💾 Size: 27 bytes

I am very interested in the memory issue, since many projects become large and consume all my session tokens and when they are close to finish I would like to be able to pause the development and continue it in my next session and I had understood that with the persistent memory it would be easier to resume the progress made.

I'm using Ubuntu 22.04 with VS Code Server.

[dylansproule]

I believe this is related:
#530

[roman-rr]

The memory tools are one of the few parts of ruflo that actually work. Your issue is likely one of these:

1. Namespace mismatch — HNSW search is per-namespace. If you stored with one namespace and search without specifying it (or with a different one), you'll get nothing back:

# Store
memory_store key='my-pattern' value='...' namespace='project'

# Search — MUST specify same namespace
memory_search query='my pattern' namespace='project'

# This returns empty (cross-namespace search doesn't work):
memory_search query='my pattern'

2. Hive-mind memory vs MCP memory — these are separate systems:

• hive-mind spawn uses its own state in .claude-flow/hive-mind/state.json
• MCP memory_store/search uses .swarm/hnsw.index + .swarm/memory.db
• They don't share data

3. Memory stats may report differently — the memory_stats tool reads from the HNSW index, but if entries were stored via the auto-memory hook they land in auto-memory-store.json (a separate JSON file with no vector search).

The memory/HNSW system genuinely works — 384-dim all-MiniLM-L6-v2 embeddings with real cosine similarity search. Just make sure you're always specifying the namespace.

For reference on what works vs what doesn't: Independent Audit