Problem or Use Case
Users who habitually start fresh sessions with /new every 1–3 turns will find that the memory nudge and skill nudge background review mechanisms never trigger.
Root cause: both counters (_turns_since_memory, _iters_since_skill) are instance variables on AIAgent, initialized to 0 in __init__. /new creates a brand-new AIAgent instance → counters reset to zero.
Impact:
- Memory nudge: requires 10 user turns within the same session (
memory.nudge_interval defaults to 10). Frequent /new users never reach 10.
- Skill nudge: requires 10 tool-calling iterations within the same session (
skills.creation_nudge_interval defaults to 10). Likewise never reached.
System prompt guidance (MEMORY_GUIDANCE, SKILLS_GUIDANCE) still exists and can prompt the model to proactively use these tools, but the post-turn "review the conversation above and consider saving" prompt never fires, so the practical effect is severely reduced.
Result: for heavy /new users, the self-improvement mechanisms are effectively dead — the memory store barely grows, and the skill library depends entirely on manual maintenance.
Proposed Solution
This is an open discussion. Known facts and constraints for reference:
Should nudge state persist across sessions? If so, what mechanism (session DB, file-based counter, time-decay)? Should /new itself trigger a review before discarding the old session? Could the nudge be time-based in addition to turn-based?
Would love to hear more use cases and ideas.
Alternatives Considered
(Left open for discussion)
Feature Type
Performance / reliability
Scope
Medium (few files, < 300 lines)
Contribution
Problem or Use Case
Users who habitually start fresh sessions with
/newevery 1–3 turns will find that the memory nudge and skill nudge background review mechanisms never trigger.Root cause: both counters (
_turns_since_memory,_iters_since_skill) are instance variables onAIAgent, initialized to 0 in__init__./newcreates a brand-newAIAgentinstance → counters reset to zero.Impact:
memory.nudge_intervaldefaults to 10). Frequent/newusers never reach 10.skills.creation_nudge_intervaldefaults to 10). Likewise never reached.System prompt guidance (
MEMORY_GUIDANCE,SKILLS_GUIDANCE) still exists and can prompt the model to proactively use these tools, but the post-turn "review the conversation above and consider saving" prompt never fires, so the practical effect is severely reduced.Result: for heavy
/newusers, the self-improvement mechanisms are effectively dead — the memory store barely grows, and the skill library depends entirely on manual maintenance.Proposed Solution
This is an open discussion. Known facts and constraints for reference:
AIAgent.__init__(run_agent.py:1688-1689)memory.nudge_intervalandskills.creation_nudge_intervalexist_spawn_background_review()afterrun_conversation()returnsShould nudge state persist across sessions? If so, what mechanism (session DB, file-based counter, time-decay)? Should
/newitself trigger a review before discarding the old session? Could the nudge be time-based in addition to turn-based?Would love to hear more use cases and ideas.
Alternatives Considered
(Left open for discussion)
Feature Type
Performance / reliability
Scope
Medium (few files, < 300 lines)
Contribution