feat: long-term agent memory across sessions (sidecar storage)

[rosspeili]

Problem Statement

Agents currently start every session with zero context of past interactions. Even if you speak with the same "Elena" or "Viktor" persona across multiple days, they have no memory of your previous debates, preferences, or established facts.

Proposed Solution

1. Implement a local JSON or SQLite "sidecar" file for each unique agent persona.
2. At the end of a session, trigger a "memory consolidation" step where the agent (via the orchestrator or a final self-summary turn) distills key facts or user preferences from the transcript.
3. On the next session start, inject this distilled memory into the agent's system prompt or as a specialized "recollection" message in the history.

Alternatives Considered

Vector database persistence, which may be overweight for simple local-first persona memory.

Priority

Low 🟢

Additional Context

This enables evolving, "living" personas that feel like true long-term collaborators rather than stateless functions.