Skip to content

feat(skills): add hindsight-memory skill for persistent agent memory#372

Closed
benfrank241 wants to merge 2 commits intoNVIDIA:mainfrom
benfrank241:feat/hindsight-memory-skill
Closed

feat(skills): add hindsight-memory skill for persistent agent memory#372
benfrank241 wants to merge 2 commits intoNVIDIA:mainfrom
benfrank241:feat/hindsight-memory-skill

Conversation

@benfrank241
Copy link

Summary

  • Adds a new hindsight-memory agent skill that teaches sandboxed agents how to use Hindsight for persistent memory across ephemeral sandbox sessions
  • Includes 5 workflows: recall before work, retain after work, reflect for synthesis, retain-files for bulk import, and cross-sandbox continuity via session bookends
  • Ships with a CLI reference, example network policy (read-write to Hindsight API), and covers both cloud and self-hosted deployments
  • Updates CONTRIBUTING.md skills table with the new entry under Platform

Closes #371

Changes

File What
.agents/skills/hindsight-memory/SKILL.md Skill definition with setup, 5 workflows, policy examples, companion skills table
.agents/skills/hindsight-memory/cli-reference.md Hindsight CLI command reference with flags and API endpoints
.agents/skills/hindsight-memory/example-policy.yaml Complete sandbox policy template for Hindsight API access
CONTRIBUTING.md Added hindsight-memory to the skills inventory table

Test plan

  • Skill files are valid markdown with correct YAML frontmatter
  • example-policy.yaml validates against the policy schema in architecture/security-policy.md
  • Skill is discoverable by agent harnesses via .agents/skills/ directory convention
  • CLI commands in the skill and reference match the current Hindsight CLI

🤖 Generated with Claude Code

Adds a new agent skill that teaches sandboxed agents how to use Hindsight
for persistent memory across ephemeral sandbox sessions. Includes the skill
definition with recall/retain/reflect workflows, a CLI reference, and an
example network policy for Hindsight API access.

Closes NVIDIA#371

Signed-off-by: Ben <ben.bartholomew@vectorize.io>
@github-actions
Copy link

github-actions bot commented Mar 16, 2026

All contributors have signed the DCO ✍️ ✅
Posted by the DCO Assistant Lite bot.

@benfrank241
Copy link
Author

I have read the DCO document and I hereby sign the DCO.

@benfrank241
Copy link
Author

recheck

… skill

- Fix --budget values: "medium" → "mid" (matches CLI default)
- Fix --fact-type values: "observation" → "opinion" (matches CLI)
- Fix --context description: it's freeform text, not a structured tag
- Fix API endpoints table: add missing POST /memories (retain),
  fix documents endpoint to GET (list), not POST
- Replace deprecated `bank background` with `bank mission`
- Fix config setup: use `hindsight configure` instead of writing
  TOML with shell variable interpolation (which doesn't work)
- Add missing reflect flags: --max-tokens, --schema
- Add missing recall flags: --include-chunks, --chunk-max-tokens

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@johntmyers
Copy link
Collaborator

This does not conform to our contributing requirements and is the PR is functionally un-sound.

@johntmyers johntmyers closed this Mar 16, 2026
@benfrank241
Copy link
Author

Thanks for the review @johntmyers. Could you share specific details on what doesn't conform or what's functionally unsound? Happy to fix — just want to make sure I address the right things.

A few areas I'm already aware could use improvement:

  • The --config HINDSIGHT_API_URL in the provider create example uses a bare key without =VALUE, which may not be valid for --config (it is for --credential)
  • The PR body used a custom format rather than the PR template sections

If there are other issues (policy format, command syntax, style), I'd appreciate knowing so I can get them all in one pass.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

wontfix This will not be worked on

Projects

None yet

Development

Successfully merging this pull request may close these issues.

feat(skills): add hindsight-memory skill for persistent agent memory across sandbox sessions

2 participants