New: Affective Goal Drive + Self-Adjudicating Memory (PLAN-22 and PLAN-23)

[VGIL77]

Two memory and autonomy upgrades just landed on main. Both are on by default and both are designed to be safe and reversible, so here is what they do and what you can expect.

1. Affective Goal Drive: the agent decides when work deserves a plan (PLAN-22)

Bitterbot now appraises the complexity of each request as it arrives and, when the work is genuinely multi-step, automatically spins it up as a goal-oriented task with checkpoints instead of trying to one-shot it.

What you can expect:

• Trivial and normal requests are untouched. A fast, local heuristic scores every prompt; the model is only consulted for genuinely ambiguous cases, so there is no added latency or cost on ordinary turns.
• Big requests become durable tasks. "Refactor across all three packages, add tests, and open a PR" becomes a tracked goal that survives context limits and restarts, rather than dying halfway.
• It is affect-aware. The threshold for escalating to a goal is modulated by the agent's hormonal state: more conservative under stress, more eager when rewarded. This is the part no off-the-shelf agent does.
• You are always in control. When a task is auto-initiated, the agent tells you and offers an immediate opt-out for that turn (reply "just answer" to get a plain inline response). To turn the behavior off entirely, set BITTERBOT_TASKS_AUTO_INITIATE=0, or BITTERBOT_TASKS_COMPLEXITY_GATE=0 to disable appraisal too.

2. Self-Adjudicating Bitemporal Memory: beliefs that can be questioned and revised (PLAN-23)

Bitterbot's knowledge graph now tracks relationships as a versioned belief history instead of a flat set of facts. When two things it learns contradict each other, it notices, keeps both, and resolves the conflict carefully during a rest cycle rather than overwriting blindly.

What you can expect:

• Contradictions are flagged, never silently merged. If the agent learns conflicting facts, both are kept active and the conflict is recorded for later adjudication.
• Nothing is ever deleted on the spot. The write path is strictly non-destructive. The only place a belief is closed is during a "reconsolidation" dream, and only after the supporting evidence has settled and a confidence threshold is cleared. That threshold is also affect-gated: under stress, the agent is more conservative about revising what it believes.
• History is queryable. Superseded beliefs are retained with their validity intervals, so the system can answer "what did I believe about X, and when did that change."
• On by default, reversible. Set BITTERBOT_KG_RELATIONSHIPS=0 to disable relationship population if you prefer the previous behavior.

Design notes

Both features build directly on the existing long-horizon task spine and the SAGE graph-memory substrate rather than bolting on new systems. The guiding principle throughout: the write path stays cheap and non-destructive, and anything irreversible happens later, during rest, gated by the agent's own affective state. Full design docs live under docs/plans/ (PLAN-22 and PLAN-23) and the memory behavior is documented in docs/memory/sabm-belief-adjudication.md.

Feedback welcome. If you try the goal workflows or notice the memory revising a belief, we would love to hear how it behaves for you.