Local-first 🧠memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.
Self-Learning Skill Proposals - implementation plan
Phase 0 outcome
Recommended architecture: file-first, dream-adjacent modules for proposal learning, procedural memory, and recurrence support. The modules share deterministic helpers but do not merge into one engine and do not modify the existing dream preference path.
Phase 2 TDD plan
Brain path and artifact contracts
- Add tests for proposal/procedure/procedural-index/recurrence paths staying inside
Brain/and rejecting unsafe slugs. - Implement path helpers and any narrow shared types.
- Commit:
feat(brain): add procedural learning artifact paths.
- Add tests for proposal/procedure/procedural-index/recurrence paths staying inside
Skill proposal storage and detectors
- Add failing core tests for repeated-action, structural-similarity, co-occurrence, temporal-routine, duplicate suppression, and watermark advancement.
- Implement deterministic proposal scan over Brain log/session events with conservative detector rules and stable slug/hash behavior.
- Commit:
feat(brain): detect skill proposal patterns.
Proposal review lifecycle
- Add failing core and CLI tests for list/show/accept/reject, accepted procedure artifact creation, rejected proposal suppression, audit metadata, and JSON output.
- Implement proposal review core and CLI verbs.
- Commit:
feat(brain): add skill proposal review workflow.
Procedural memory reconciler
- Add failing tests for frontmatter parsing, scanning Brain procedures and installed skill/runbook roots, stale deletion, source-change updates, and stable IDs.
- Implement procedural index and CLI list/reconcile behavior.
- Commit:
feat(brain): index procedural memory entries.
Procedural usage sidecar
- Add failing tests proving usage updates do not rewrite source
SKILL.md/procedure files and that last-used/count metadata appears in list/export output. - Implement usage sidecar and CLI mark-used behavior.
- Commit:
feat(brain): track procedural memory usage.
- Add failing tests proving usage updates do not rewrite source
Recurrence support ledger
- Add failing tests for same-scope duplicate support increment, cross-scope recurrence evidence, thresholded diagnostic promotion state, locked/no-auto-change behavior, reference-counted forget, and source purge.
- Implement recurrence ledger and diagnostics without mutating preferences or auto-promoting scope.
- Commit:
feat(brain): record procedural recurrence support.
MCP and integration surfaces
- Add focused MCP tests for read/diagnostic proposal/procedural surfaces if exposed.
- Implement handlers only for surfaces that remove meaningful CLI friction.
- Commit:
feat(mcp): expose procedural learning diagnosticsif needed.
Docs and versioning
- Update CLI reference, MCP docs if applicable, README feature summary, CHANGELOG, and version files.
- Run formatter/linter before commit as requested.
- Commit:
docs: document procedural learning workflowor version-specific release commit.
Validation checklist
- Focused tests after each TDD unit.
bun run fmtandbun run lintbefore every commit.- Full QA before pre-push stop point:
bun run test,bun run typecheck,bun run lint, version sync check if available, andcode_checker. - Self-review compares branch diff against
mainand checks the implementation against the selected task acceptance criteria.
Stop points
- Phase 6 pre-push chat stop-point with Russian
ask_reportbefore pushing. - Release/pre-release stop-point per feature-release-playbook after PR approval/merge flow.