Memory Moments and Curiosity Threads

Moments: Semantic Bookmarking

Introduced “Moments,” a feature allowing users to pin Chalie’s responses as enriched bookmarks. Unlike simple history, Moments are indexed via pgvector for instant semantic recall. A background worker performs “enrichment” over a 4-hour window, gathering surrounding conversation gists to contextualize why a moment was saved. Pinning also applies a salience boost to related memories, ensuring relevant past context surfaces more frequently in cognitive drift and context assembly.

From Goals to Curiosity

We have significantly pivoted the agent’s long-term drive by replacing the rigid goal_service with a fluid Curiosity Pursuit system. Instead of tracking static tasks, Chalie now generates “Curiosity Threads”—self-defined interests that manifest as impromptu human interactions or autonomous research. This includes a new engagement tracker and seed action logic to help Chalie decide when to pursue a thread based on context relevance and “surprise” factors.

Scheduler Refactor

Refactored the scheduler to move away from the ambiguous “reminder vs. task” distinction. We now use a strict notification vs. prompt split. Notifications surface text directly to the interface on a timer (bypassing the LLM), while Prompts are fed back into the LLM pipeline as a conversational turn. This makes the intended behavior predictable for both the user and the model.

Refined Intent Execution and UI

Fixed several bugs in the tool_worker affecting multi-step skill calling, specifically improving the model’s ability to reason from its own knowledge if research tools return empty results. On the frontend, we added a 10-second “undo” window for forgetting moments to prevent accidental data loss. The moment cards also received a visual overhaul with better markdown rendering and collapsible message blocks to reduce visual repetition between the pinned text and the generated summary.