strata-memory

Strata Memory

Purpose

strata-memory is an open-source MCP memory server (v2.1.0, MIT) for one agent or several agents sharing a palace.

• SQLite is the single source of truth (memories, audit, trajectories).
• Chroma is a companion vector index. Rebuild it from SQLite; do not treat it as authoritative.
• Markdown from strata_project is a read-only human view.
• Writes go through a Quality Kernel and CBT middleware.
• Recall is two-step: recall_context returns short cards (id, summary, score). expand_memory_detail loads full text for one id.

Repository: https://github.com/SherryLi-vc/strata-memory

This skill does not bundle the server. Install it from that repository. The strata-memory-mcp package is not on PyPI, so do not use uvx.

When to Use

• The user wants persistent memory, long-term context, or an MCP memory server
• Several agents should share durable facts through one palace
• Session scratch notes should become durable only after the user confirms them
• Recall should stay small: search cards first, then expand the hits that matter

When not to use

• Storing passwords, API keys, tokens, or other secrets
• Recording relative times (yesterday, 刚才, today) inside fact_claim; use an ISO date
• Storing hedged wording (probably, maybe, 也许) as memory_type: factual_truth
• Rebuilding the vector index unless the user has agreed (confirm=true)

Data and network

Tell the user this before the first write or recall:

• Rows live under STRATA_PALACE (default ~/.strata/palace).
• The configured embedding endpoint receives memory text and recall queries (commit_memory, promote_session, recall_context, and index rebuild), unless the provider is local.
• Default option is SiliconFlow: provider siliconflow, model BAAI/bge-m3, base URL https://api.siliconflow.cn/v1. That model returns 1024 dimensions and is truncated to STRATA_EMBEDDING_DIMENSION (default 384).
• STRATA_EMBEDDING_PROVIDER=openai or openai_compatible calls any OpenAI-compatible POST /embeddings endpoint. Set STRATA_EMBEDDING_BASE_URL and STRATA_EMBEDDING_MODEL. The vector length must equal STRATA_EMBEDDING_DIMENSION (OpenAI text-embedding-3-small uses 1536 when that provider is selected). Then rebuild the index.
• STRATA_EMBEDDING_PROVIDER=local runs an in-process model (default sentence-transformers/all-MiniLM-L6-v2, 384 dimensions). Install with uv sync --extra local. Text stays on the machine.
• A local server (Ollama, llama.cpp, TEI, vLLM) is openai_compatible with a loopback STRATA_EMBEDDING_BASE_URL such as http://127.0.0.1:11434/v1. Loopback does not require an API key. Text is sent to that URL.
• Key lookup for HTTP providers: explicit argument, STRATA_EMBEDDING_API_KEY, STRATA_API_KEY, then SILICONFLOW_API_KEY (SiliconFlow) or OPENAI_API_KEY (OpenAI-compatible). Empty values, values shorter than 20 characters, and values containing ... are rejected.
• Never ask the user to paste a key into chat. Never echo a key, write one into memory, or commit one. The user sets it in their own shell profile or host config.