OpenPocket

Appearance

Sessions and Memory ​

OpenPocket persists execution and context as markdown/json artifacts under the selected agent workspace.

In multi-agent installs, every agent has its own workspace and memory timeline. Sessions and memory are not shared across agents.

Workspace Bootstrap ​

On config load / onboarding, ensureWorkspaceBootstrap ensures directories:

  • workspace/memory
  • workspace/sessions
  • workspace/skills
  • workspace/scripts
  • workspace/scripts/runs
  • workspace/cron
  • workspace/.openpocket

And default files (if missing):

  • AGENTS.md
  • BOOTSTRAP.md (fresh workspace onboarding gate)
  • SOUL.md
  • USER.md
  • IDENTITY.md
  • TOOLS.md
  • HEARTBEAT.md
  • MEMORY.md
  • PROFILE_ONBOARDING.json
  • TASK_PROGRESS_REPORTER.md
  • TASK_OUTCOME_REPORTER.md
  • BARE_SESSION_RESET_PROMPT.md

This happens independently per agent workspace.

Onboarding Lifecycle Persistence ​

Workspace state file:

  • workspace/.openpocket/workspace-state.json

Key timestamps:

  • bootstrapSeededAt
  • onboardingCompletedAt

When onboarding completes:

  • IDENTITY.md and USER.md are persisted
  • BOOTSTRAP.md is removed
  • onboardingCompletedAt is written

Session Lifecycle ​

For each task, runtime creates:

  • workspace/sessions/session-<timestamp>.md

During run, each step appends:

  • thought text
  • normalized action JSON
  • executor result text

At completion/failure, runtime appends final block with status and message.

Daily Memory ​

Runtime appends one compact line per task to:

  • workspace/memory/YYYY-MM-DD.md

Line contains:

  • local time
  • status (OK or FAIL)
  • model profile key
  • task summary
  • compact result summary

Memory Recall Tools ​

Task loop includes memory-specific tools:

  • memory_search: search MEMORY.md + memory/*.md
  • memory_get: read targeted line ranges from those files

Prompt policy encourages memory-first retrieval before answering prior-decision/preference/history questions.

Screenshots ​

When enabled (screenshots.saveStepScreenshots=true):

  • each step screenshot is saved locally in the selected agent's state/screenshots/
  • optional screenshot path is appended to step result text
  • debug marker overlays may be saved for tap/swipe steps

Retention policy:

  • keep newest screenshots.maxCount
  • delete oldest PNG files when over limit

Auto Task Sedimentation ​

After successful tasks, runtime may auto-generate reusable artifacts:

  • skill markdown: workspace/skills/auto/*.md
  • replay script: workspace/scripts/auto/*.sh

These artifacts are surfaced in task result and used by outcome narration.

Current auto-skill experience behavior:

  • generated skills include a behavior_fingerprint for replay dedupe
  • step-derived ui_target semantic traces are embedded for element intent reuse
  • runtime injects summary metadata only; the model must read(location) to load a skill body when it decides one is relevant

This experience layer is agent-local. A second agent does not see or reuse the first agent's auto-generated memory unless you manually copy files between workspaces.