Add memory storage engine

[theantichris]

Summary

Add the core memory storage engine that handles Markdown file I/O, SQLite indexing, text chunking, and hybrid search (vector + BM25).

Details

New dependency:

• modernc.org/sqlite — pure Go SQLite driver (no CGo, preserves clean cross-compilation)

New files:

• internal/memory/store.go
• internal/memory/store_test.go

Key types:

type EmbedFunc func(ctx context.Context, text string) ([]float64, error)

type Store struct {
    db      *sql.DB
    dir     string       // ~/.config/ghost/memory
    embedFn EmbedFunc
    mu      sync.RWMutex
}

type SearchResult struct {
    Source  string  // e.g. "daily/2026-02-07.md" or "MEMORY.md"
    Content string
    Score   float64
}

SQLite schema:

• chunks table: id, source, chunk_index, content, embedding (BLOB)
• chunks_fts FTS5 virtual table with sync triggers
• WAL mode + busy_timeout for concurrency

Functions:

• NewStore(dir, embedFn) — create directory structure, open DB, create schema
• SaveToDaily(ctx, content) — append to daily/YYYY-MM-DD.md, index new content
• SaveToLongTerm(ctx, content) — append to MEMORY.md, index new content
• IndexFile(ctx, source) — chunk file, embed chunks, upsert into SQLite
• Search(ctx, query, topK) — hybrid retrieval (70% cosine similarity + 30% BM25)
• chunkText(text, maxChars, overlapChars) — paragraph-aware chunking (~400 tokens, 80 overlap)
• cosineSimilarity(a, b []float64) — vector math helper
• Close() — close SQLite connection

File structure on disk:

~/.config/ghost/memory/
  MEMORY.md
  daily/
    2026-02-07.md
  index.db

Design decisions:

• EmbedFunc injection (not an interface) for testability
• Brute-force vector search (corpus is small enough, no ANN needed)
• Markdown files are source of truth; SQLite index is derived and rebuildable
• Embedding vectors serialized as binary (compact, fast)

Testing:

• Tests use temp dirs + mock EmbedFunc returning deterministic vectors
• TestChunkText, TestCosineSimilarity, TestNewStore, TestSaveToDaily, TestSaveToLongTerm, TestSearch, TestIndexFile

Dependencies

• Add Ollama embedding client #235 (Ollama embedding client — for EmbedFunc type contract, though tests mock it)

Acceptance Criteria

• Directory structure created on NewStore()
• SQLite schema with FTS5 created correctly
• SaveToDaily creates/appends daily Markdown files
• SaveToLongTerm creates/appends MEMORY.md
• IndexFile chunks text and stores embeddings
• Search returns relevant results with hybrid scoring
• All tests pass with mock embedding function

[theantichris]

Storage Path: Use XDG Data Dir

The default storage path should be changed from ~/.config/ghost/memory/ to ~/.local/share/ghost/memory/ to follow XDG Base Directory conventions.

Memory data (facts, summaries, embeddings) is user data, not configuration. Per the XDG spec:

• ~/.config/ ($XDG_CONFIG_HOME) → configuration files (e.g., config.toml)
• ~/.local/share/ ($XDG_DATA_HOME) → user data files

This also aligns with conversations storage (#195), which should use ~/.local/share/ghost/conversations/.