AgentSkillsCN

Memory Vector

内存向量

SKILL.md

memory-vector

Semantic vector search for agent memory files using local embeddings (100% free).

Stack

  • •Vector DB: LanceDB (serverless, file-based)
  • •Embeddings: Ollama + nomic-embed-text (local, free)
  • •Cost: $0

Installation

Requires Ollama with nomic-embed-text model:

bash
ollama pull nomic-embed-text

Usage

Index all memory files

bash
cd ~/clawd/skills/memory-vector
node index.js --index

Indexes:

  • •memory/*.md (daily logs)
  • •MEMORY.md (curated memory)
  • •SESSION-STATE.md (active task)

Search for similar content

bash
node index.js --search "model routing decisions"
node index.js --search "what did we learn about S3"
node index.js --search "errors with authentication"

Show index stats

bash
node index.js --stats

How It Works

  1. •Chunking: Splits files by ## headers (or paragraphs if no headers)
  2. •Embedding: Calls Ollama API for nomic-embed-text vectors (768 dimensions)
  3. •Storage: LanceDB stores vectors in ~/.lancedb/ (file-based, no server)
  4. •Search: Approximate nearest neighbor search with cosine similarity

Comparison to memory_search

Featurememory_searchmemory-vector
Search typeKeyword + basic semanticTrue vector similarity
"Find similar to X"❌ Limited✅ Native
"That AWS issue" (no exact match)❌ Miss✅ Finds it
SetupZero5 min
CostFreeFree

Re-indexing

Run indexing again to update after adding new memory files:

bash
node index.js --index

Suggested: Add to heartbeat or daily cron.

Troubleshooting

Ollama not running

bash
ollama serve &
# or
systemctl start ollama

No embeddings model

bash
ollama pull nomic-embed-text

LanceDB errors

Delete and re-index:

bash
rm -rf ~/.lancedb
node index.js --index