version: "1.0.0" name: experience-memory description: Save successful action tips as soon as they work and retrieve them for later actions under matching observable conditions; distill reusable procedures into learned skills with evidence. Use while controlling the robot, when building automation, and at the start of work or after compaction.

Experience memory

memory/ belongs to this task and persists across episodes, runs and context compaction. Keep learned procedures under memory/skills/, apart from the provided skills.

Retrieve

At the start of work or after compaction, read memory/INDEX.md and memory/TIPS.md if present, and open only the learned skills whose conditions match the current scene. Open linked evidence only when applicability or outcome is uncertain; do not load the whole history. Current perception and the live tool schemas take precedence over anything remembered.

Record tips immediately

• As soon as an action produces a visibly successful, useful result, add or update one

row in memory/TIPS.md. A visible local result is enough; task-level confirmation can be added later.

• Record the task phase, the observable preconditions, the tool and arguments or the

reusable pose relation, the visible effect, and the saved frame paths the tools returned. Say what made it work, so later attempts skip the exploration and the superseded failures.

• Refine the matching row when the same maneuver improves instead of adding variants.
• Scenes drift between attempts (objects moved back by hand, lighting, layouts): record

the conditions under which a tip held, not only its numbers.

Distill

Promote a tip to memory/skills/<slug>/SKILL.md plus evidence.json when it shows a reusable motion, contact strategy, recovery or completed procedure. Keep the procedure short and relative to observed features; keep attempt-specific values and evidence links in evidence.json. Encode poses as the robot's pose-record conventions define them, when robot-interface has them.

Update memory/INDEX.md with the skill path, tags, confirmation level and the last episode or date. Mark task-confirmed only when the harness's trusted check accepted an attempt the procedure contributed to, or every TASK.md condition was checked in fresh images; otherwise observed. Revise a skill when later evidence contradicts it, and keep its provenance.

Read references/format.md when creating or updating a learned skill. When a procedure has worked unattended and a persistent library exists, also contribute it there (continual-learning).

Code that runs without memory

A controller run for evaluation may not see the editable workspace: put every procedure it needs into its code before freezing it. When you continue from carried notes on a new scene, evidence paths may no longer resolve: rely on the stated conditions and pose recipes, replay relative recipes from the current measured state, and reacquire every point a recipe anchors to.