Full-text recall over raw conversation history.

Distinct from semantic memory (Raxol.Agent.Memory.Store.Ets, which recalls curated Records), this is an inverted index over Raxol.Agent.Conversation item text. It answers "what did we say about X earlier" by returning the RAW matching items, not summaries.

Items are tokenized with Raxol.Agent.Memory.Record.tokenize/1 (the same tokenizer the semantic store uses) and ranked with the same length-normalized BM25-lite scoring. The index is fed either explicitly with index/2, or by attach/3, which subscribes to a Raxol.Agent.Conversation.Log and indexes its snapshot plus every appended item.

Tables (derived from the registered name)

• primary :set { {conversation_id, seq}, %{item, len} }
• Tok :bag {token, {conversation_id, seq}} -- inverted index
• DF :set {token, doc_freq} plus {:__N__, doc_count} and {:__DL__, total_doc_len} for idf and average document length