Skill Sedimentation

What It Solves

Agents often finish non-trivial tasks while useful context disappears: an error is fixed but the root cause is not captured, a workflow succeeds but the Skill is not updated, and user preferences are re-explained later. Skill Sedimentation treats post-task learning as a controlled knowledge-write pipeline rather than free-form memory.

How It Works

• Trigger / Retrieve: after task completion or a significant correction, it reads experience files under _core_files/memory/, treating missing files as empty lists.
• Reflect: it appends one scored, tagged entry to reflections.md, keeping the loop concise instead of dumping broad conclusions.
• Distill: it routes findings by type—error causes to TABOOS.md, workflow improvements to _core_files/skills/, project-specific state to dated memory files, and general principles to MEMORY.md.
• Promote: it checks an adaptive priority threshold before turning repeated experience into durable rules or new Skill files, with frequency controls and manual override lockout.

Boundaries

• Writes are limited to _core_files/memory/ and _core_files/skills/, and path checks must run before amplitude checks.
• Promotions that add capability domains, rewrite core rules, or add external dependencies are rejected by default pending human confirmation.
• Best for local agent memory systems; it is not a cross-account knowledge store or an unattended mechanism for changing system configuration.