abd-search-memory

name	abd-search-memory
description	Semantic search over a local FAISS vector index built from document chunks. Use when the user says "use memory", "search memory", "what does memory say", "from our content", "what do we have on [topic]", or asks about content that has been ingested into agent memory.
license	MIT
metadata	{"author":"agilebydesign","version":"1.0.0"}

Search Memory

Run semantic search against a FAISS index and inject the best-matching chunks into the conversation. Matches by meaning, not only keywords.

How to search

python scripts/search_memory.py "<query>" --rag <source>/memory/rag --k 5

• --rag: Path to the rag/ folder (default: <ROOT>/memory/rag)
• --k: Number of results (default 5)
• --format: text (default) or json

When to activate

Run search_memory.py when the user says:

• "use memory", "search memory", "what does memory say"
• "from our content", "from our decks", "from ABD materials"
• "what do we have on [topic]"
• Asks about any topic that has been ingested into memory

Agent flow

1. Derive query — Extract a semantic query from the user's question
2. Run search — Execute search_memory.py with the query against the corpus index
3. Inject results — Use the returned chunks in your response
4. Cite sources — Include path and slide/section when using retrieved content. If chunks have front matter, note evidence_type / chunk_type to signal weight

Active memory for this conversation

When the user points at a corpus — by naming a topic folder, having CONTENT_MEMORY_ROOT in conf/.secrets, passing --rag, or saying "use this memory" — treat that memory/rag/ as the active retrieval source for the rest of this conversation. For later questions, run semantic search against that same index by default until the user explicitly changes or clears it (e.g. "use a different memory folder", "stop using that corpus"). Do not make them repeat the path on every turn. Workspace policy: AGENTS.md.

Chunk front matter (when spec was active during chunking)

Results may include chunks with YAML front matter. Use the labels to calibrate your response:

chunk_type	How to use
Normative roles (e.g. stat_block, rules-like sections)	Cite carefully — check against source
Illustration / example scenes	Use as illustration unless corroborated
toc_or_index, noise_or_art, navigation-like	Low weight

Configuration

• OPENAI_API_KEY must be set (for query embedding)
• Index must exist: run embedding first (see abd-embed-vectors)
• Dependencies: pip install -r requirements-rag.txt

Gotchas

• If --rag is omitted, search uses <ROOT>/memory/rag where ROOT comes from CONTENT_MEMORY_ROOT (usually set in conf/.secrets) or cwd.
• Search requires the same embedding model used to build the index.
• An empty or missing index returns no results — run the embed skill first.