What does this skill do, and when should you use it?
This skill helps coding agents preserve and recover working context during long-running work, disconnects, VS Code restarts, branch switches, and handoffs. It uses a session directory in the repository and maintains human-readable state, timeline, file, and handoff records. The files are intended to let another agent resume with minimal context loss. It is the NeMo-RL session-memory skill in NVIDIA’s skills catalog and is licensed Apache-2.0.
Creates or reuses a repository-local session/<timestamp> directory; generates the directory name from local machine time; writes or updates session_state.md, timeline.md, files.md, and handoff.md; records goals, subtasks, loaded skills, status, plans, assumptions, blockers, major actions, command results, inspected files, and next steps; on recovery, reads the handoff, state, and recent timeline, then performs lightweight git-state verification.
- A coding agent making nontrivial edits that needs a checkpoint before pausing or risking disconnection.
- A developer restarting VS Code or switching branches while preserving active task context.
- A team handing off a long-running experiment or unfinished development task between agents.
- A user resuming work in a repository that already contains session/ directories.
How do you install this skill?
- Session files may contain source code, paths, decisions, or other sensitive information; the skill defines no redaction, retention, access-control, or encryption policy.
- The mkdir and timestamp workflow does not specify handling for concurrent sessions, name collisions, permission failures, or partial writes.
- The benchmark pass rates and version information lack independent corroboration in the supplied material and should not be treated as executed verification.
- Chinese-language documentation and interaction requirements are not specified, so Chinese users may need to interpret the fields and recovery workflow themselves.
- Shell / CLI
- Local filesystem
Install it from the NVIDIA/skills collection with the skills CLI: npx skills add nvidia/skills --skill nemo-rl-session-memory --yes. The CLI prompts for the skill and installation destination; the source does not specify a fixed client-specific folder path.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/nemo-rl-session-memory" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Create a session checkpoint for this coding task before we pause, and record the next actions.
- Read the existing session/ records and restore the current working context.
Invoke it for a relevant task with a request such as: “Create a session checkpoint for this coding task before we pause, and record the next actions.” To recover, ask: “Read the existing session/ records and restore the current working context.” Do not use it for simple questions, short tasks, one-off commands, linting, or code review.
What are this skill's strengths and limitations?
- Covers explicit recovery scenarios including disconnects, restarts, handoffs, and branch switches.
- Produces structured, human-readable Markdown records.
- Separates state, timeline, file tracking, and handoff guidance for easier resumption.
- Requires repository filesystem access and basic shell command execution.
- The source provides no automatic cleanup for old sessions, conflict handling, or test suite.
- It manages working context only; it does not preserve complete conversation history or replace code and git state.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| NeMo-RL Session Memory this page ✓ NVIDIA · Official | 51 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| LeanCTX — Local Context Intelligence Layer | 61 · Recommended | ★ 3.9k | 13d ago | Apache-2.0 |
| neat-freak — Knowledge & Governance Closeout | 63 · Recommended | ★ 21k | 26d ago | MIT |
| Obsidian Knowledge Brain v4.0 | 56 · Use with care | ★ 100 | 3mo ago | MIT |
| Persistent File Planning | 49 · Use with care | ★ 27k | 13d ago | MIT |
How did FollowSkills review this skill?
The skill limits writes to a repository-local session/ directory and requests human-readable records; no external network access, credential access, destructive action, or covert exfiltration is shown. It lacks sensitive-data exclusion rules, user confirmation, explicit permission boundaries, rollback guidance, and data-flow disclosure, so points are deducted.
The directory/file workflow, recovery order, and git-state verification are clear. However, it relies on shell commands and local time without diagnostics for collisions, permission failures, corrupted sessions, or other abnormal conditions, and no tests cover this skill's key paths; the static cap keeps the score at or below 10 with further deductions.
Use cases, activation conditions, and exclusions are reasonably explicit for disconnects, restarts, handoffs, and long tasks. Input/output contracts, non-code-repository boundaries, Chinese-language guidance, and more precise semantic trigger conditions are not defined, so points are deducted.
SKILL.md provides name, description, license, use cases, and a basic layered structure. The skill card adds owner, version, risks, and evaluation summary, but Instructions, Examples, FAQs, author/tags, a clear changelog, and an explicit maintenance/update path are missing; the stated version is not directly evidenced by the supplied files, so points are deducted.
The expected files, checkpoint cadence, and recovery workflow are directly usable for common session-memory tasks, and the evaluation file reports successful positive-task behavior. Static materials cannot verify actual file creation, cross-environment recovery, or marginal benefit over manual notes; execution correctness remains dependent on the agent, so the score stays below the static cap.
The repository includes an evaluation report, positive and negative routing cases, and expected behaviors, providing limited auditability. The report is supplied by the repository rather than independently corroborated, and no skill-specific CI or committed tests cover the key paths, so the score is capped at 5 and reduced.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →