Skill auto-versioning
Skill auto-versioning is the agent's pattern of saving each variant of a learned how-to as a versioned record, with usage counters and last-used timestamps. The agent picks the most recent successful version automatically; the user can roll back if a newer version regressed.
Why it matters
Agents that learn how-tos (procedures the user wants the agent to remember) have a problem: the user keeps refining what they want. Today the user says "save my morning brief format" with one preference; in three weeks they refine the format; in three months they refine again. Each iteration is a real version of the same skill.
Without versioning, the agent has only the latest. If the latest is wrong (the user's last refinement was a misunderstanding, or they changed their mind), the previous good version is lost. With versioning, every save creates a new record; the agent uses the most recent active version and the older versions stay around for rollback.
The pattern also helps with usage tracking. A skill that has been called 200 times is a different signal from one that was saved and never used again. Auto-versioning attaches timesUsed and lastUsedAt to each version, so the agent can deprecate stale skills and surface popular ones.
The honest constraints:
- Storage cost. Versions accumulate. Old, never-rolled-back versions take space. Eviction policies matter.
- Rollback ergonomics. "Roll back to the version from October" requires a UI; without it, the user has to ask the agent to search and pick.
How Ajeris handles this
Ajeris's skill catalog auto-versions every save. Each skill has a stable name (e.g., "morning_brief_format") and a list of versions, each with creation timestamp, content, timesUsed, and lastUsedAt. The agent uses the most recent version by default.
Users can ask the agent "go back to last week's version of how I want my morning brief" and the agent finds and activates the matching version. The audit trail records the rollback.
