NEW: AI That Learns From You
Self-ImprovingSkills
Plexon learns from each session and quietly improves its skill library. No manual editing. A small curator agent saves new skills worth keeping, refines existing ones, retires the stale, and protects everything you've pinned. The AI gets smarter the more you use it.
A Library That Grows With You
Most AI assistants forget the moment a chat ends. Plexon doesn't. After every session (or whenever you ask) a small curator agent reviews what just happened: the patterns you taught it, the corrections you gave, the workflows you repeated. The valuable stuff gets saved as a reusable skill the AI consults next time. The weak stuff is left alone. The result is an assistant that genuinely gets better the more you use it.
- No manual editing: the curator drafts, refines, and retires skills for you
- Reviews session activity, identifies repeating patterns, captures the gist
- Existing skills get sharpened over time when the AI sees better examples
- Stale or contradicted skills are retired so the library stays signal-rich
- Smarter assistance compounds: month two is meaningfully better than month one
Pin What You Want to Keep
You stay in control. Pin any skill (bundled, hand-written, or curator-saved) and it's protected from automatic refinement and retirement. The curator can suggest changes, but pinned skills require your explicit approval. Everything else can be tuned freely. Pin sparingly for hard rules; leave the rest unpinned so the library can self-tune.
- One-click pin from any skill card in the Skills sidebar
- Pinned skills survive every curator pass untouched
- Curator can still suggest edits to pinned skills. You decide
- Mix pinned constants with self-tuning skills for the best of both
Two Modes: After-Session and On-Demand
By default, the curator runs quietly after a session ends, never blocking your work. When it finishes, a short summary of what it changed lands in your inbox (the Schedule panel's Inbox tab), so quiet doesn't mean invisible. Prefer to drive it manually? Open the wand-icon card in the Skills sidebar and run the curator on demand: review one session, the last week, or the entire workspace history. Either way, every change is logged.
- After-session mode: runs in the background, zero interruption
- Its summary shows up in your inbox after each run, so you always know what it did
- On-demand mode: trigger from the Skills sidebar wand-icon card
- Choose what to review: last session, last N sessions, or all history
- Curator runs as a sandboxed sub-agent with read-only chat access
Full Transparency Through the Curator History Viewer
Every action the curator takes is recorded: what it saved, what it refined, what it retired, and the source session that triggered the change. Open the curator history viewer from the Skills sidebar to scroll through every decision, see the before/after of any refinement, and undo any change you don't agree with. There's no black box.
- Chronological log of every curator action, with reasons
- Before/after preview for refinements: see exactly what changed
- One-click undo restores any previous version of a skill
- Filter by skill, by session, or by action type (saved/refined/retired)
- Source session links: jump straight to the chat that triggered the save
One-Click Settings in the Skills Sidebar
No buried menus. The Skills sidebar's wand-icon card gives you everything: turn the curator on or off, switch between automatic and manual mode, choose how often it runs, set the threshold for retiring stale skills, and pick which curator model to use. Every setting takes effect immediately. No restart, no config file, no command line.
- On/off toggle with immediate effect
- Automatic vs manual review mode
- Cadence: after every session, every N sessions, or daily
- Stale-skill threshold: how many sessions of disuse before retirement
- Pick the curator model independently from your chat model
Safe by Construction
The curator is sandboxed. It can only read session transcripts and write to the skills directory: no shell access, no file operations outside skills, no network calls, no tool dispatch beyond plexon_skill. It runs with strict turn and time budgets, prefers editing over deleting, and never touches a pinned skill without explicit approval. Quiet, scoped, reversible.
- Read-only access to chat transcripts; write-only access to skills
- No shell, no arbitrary file ops, no outbound network
- Strict per-run turn and time budgets: graceful timeouts
- Prefers editing over deleting: information loss is the worst outcome
- Pinned skills require explicit approval for any change
- Every action reversible from the history viewer
Under the Hood
Curator Agent
A scoped sub-agent that runs after sessions or on demand. Reads transcripts, drafts skill changes, and writes them through the same plexon_skill pipeline you use by hand: same validation, same encryption.
Pinned-Skill Protection
Anything you pin is immune to automatic refinement and retirement. The curator may suggest changes, but pinned skills require your explicit approval before anything is written.
Curator History Viewer
Every save, refine, retire, and pin/unpin is logged with the source session. Browse, filter, preview before/after, and undo any change you disagree with.
One-Click Settings
Wand-icon card in the Skills sidebar. Toggle on/off, switch automatic/manual, set cadence, configure the stale threshold, and pick the curator model, all in one place, no config files.
Sandboxed by Construction
Read-only access to chat transcripts, write-only access to skills, no shell, no arbitrary file ops, no outbound network. Hard turn and time budgets per run.
Compounding Quality
Skills get sharper across runs as the AI sees better examples. Stale skills disappear so the library stays signal-rich. Month two is meaningfully better than month one, without you writing a single instruction.
Pairs With Schedule
The same machinery that lets the curator run automatically after a session also powers Plexon's new Schedule panel: a sidebar where you can run any sub-agent on a cron expression or simple interval. The curator improves your library; the schedule turns any agent into a recurring job. Together, your AI gets smarter and works while you sleep.
Explore Schedule