global memories

README.md

@@ -101,6 +101,7 @@ okay
    - **Importance scoring:** Messages mentioning intent words ("plan", "remember", etc.), showing length, or emotional weight receive higher scores. When the store exceeds its cap, the lowest-importance/oldest memories are pruned. You can also call `pruneLowImportanceMemories()` manually if needed.
 
    - **Pattern-aware long-term recall:** Long-term memory is only queried when Nova detects a recall cue (`remember`, `do you know`, `we talked`, `refresh my memory`, etc.). When a cue fires, she fetches the top cosine-similar memories but only keeps the ones whose score meets `MEMORY_RECALL_SIMILARITY_THRESHOLD` (default 0.62); otherwise the conversation stays anchored on the short-term buffer and summary. This keeps memory-driven context from popping up during casual chat unless you explicitly ask for it.
+   - **Global recall:** `ENABLE_GLOBAL_MEMORIES` is true by default, so Nova now considers every user’s long-term memories (tagged with their `user_id`) whenever she runs recall logic. Set it to `false` only if you need strict per-user isolation again.  
    - **What gets embedded:** After every user→bot turn, `recordInteraction()` (see [src/memory.js](src/memory.js)) bundles the pair, scores its importance, asks OpenAI for an embedding, and stores `{ content, embedding, importance, timestamp }` inside the SQLite tables.
 - **Why so many numbers:** Cosine similarity needs raw vectors to compare new thoughts to past ones. When a fresh message arrives, `retrieveRelevantMemories()` embeds it too, calculates cosine similarity against every stored vector, adds a small importance boost, and returns the top five memories to inject into the system prompt.
 - **Memory cooldown:** `MEMORY_COOLDOWN_MS` (defaults to 180000 ms) keeps a long-term memory out of the retrieval window for a few minutes after it was just used so Nova has to pull fresh context before repeating herself, while still falling back automatically if there isn’t anything new to surface.