🦛 Hippo Memory Insights — 2026-06-07

[github-actions[bot]]

Summary

• Memories in store: 535 (479 observed, 53 inferred/semantic, 3 verified)
• New lessons learned from commits today: 0 (no fix/revert/bug commits in the last 7 days)
• New semantic patterns consolidated: 29 merged from episodic clusters
• Auto-shared to global store: 245 high-value memories
• Highest-confidence memories: 3 verified incident records (see Verified Memories below)

Top Memories Surfaced

1. [verified] mem_4afaa0965161 — Recurring incident: stale workflow .lock.yml files cause churn and CI friction; run make recompile after markdown workflow edits. (score=0.916 in CI recall)
2. [verified] mem_b3de87b76481 — Recurring incident: Codex auth failures can break agent jobs; verify auth/mode/token setup before reruns.
3. [verified] mem_e02072c9ae07 — Recurring incident: node: command not found on GPU/self-hosted paths when node runtime/tooling is missing; validate runtimes and PATH early.
4. [109d half-life] mem_0324334270d0 — Guard list_code_scanning_alerts calls in workflow prompts: always include state: open and severity: critical,high to avoid oversized MCP responses.
5. [109d half-life] mem_3c0154ff3f2d — Build and recompile lock files: make build then make recompile.
6. [95d half-life] mem_2704ae961669 — make build && make fmt — catch compile errors and formatting issues early.
7. [semantic] sem_1cfa35f0999d — PRs that fail CI immediately after opening are closed without merging — wasting the entire agent session. This is NOT optional.

Suggested Improvements

Error Patterns

• Stale .lock.yml files (3× confirmed incident): After any markdown workflow edit, agents must run make recompile. Consider adding a pre-commit hook or CI gate that detects stale lock files and fails fast.
• Codex auth failures: Codex auth can silently break agent jobs mid-run. Recommend adding an auth pre-check step (verify GITHUB_TOKEN / Copilot token) at the top of any workflow that uses the Codex engine.
• node: command not found on self-hosted runners: Validate node availability early (runtime check step) and document minimum Node.js version requirements per runner type. The embed pass requires Node 22.5+.
• Oversized MCP responses from list_code_scanning_alerts: Always pass state: open and severity: critical,high. Consider enforcing this as a linter rule on workflow prompts.

Code Quality

• JavaScript refactoring: The javascript-refactoring skill (skills/javascript-refactoring/SKILL.md) targets splitting large .cjs files. If any action scripts have grown beyond ~300 lines, apply the guide.
• Validation complexity: scratchpad/validation-refactoring.md documents when to split validators (>300 lines or 2+ distinct domains). Review pkg/workflow/compiler_yaml.go and related validation files against this threshold.
• Type patterns: Use typed string aliases (e.g., WorkflowID, StepID) rather than plain string to prevent cross-field bugs. Already documented but inconsistently applied.
• Code organisation: Prefer many smaller files grouped by functionality; add new files for new features rather than extending existing ones.

CI Health

• Formatting gate: Unformatted Go code causes immediate CI failures. The two-checkpoint strategy is well-documented but not always followed — make build && make fmt (Checkpoint 1) and make agent-report-progress (Checkpoint 2) before every PR.
• ANSI escape validation: validate-yaml job in .github/workflows/ci.yml scans all YAML for ANSI codes. The compiler (pkg/workflow/compiler_yaml.go) strips ANSI automatically — ensure all new string-building paths go through stringutil.StripANSI().
• Full test suite locally: make test takes >5 min; let CI run it. During development use make test-unit (~30s). This pattern is consolidated in semantic memory sem_04062c3abdb8.

Quick Wins

1. Add a CI gate for stale lock files — A simple diff check after make recompile would prevent the most common CI friction pattern.
2. Enforce list_code_scanning_alerts filter — Add a linter rule (or workflow schema validation) requiring state and severity on all list_code_scanning_alerts MCP calls.
3. Node runtime pre-check step — Add a node --version check early in any workflow that needs Node, with a clear error if missing.
4. Promote checkpoint docs to a top-level AGENTS.md callout — The two-checkpoint validation strategy is the highest-traffic pattern in the store (many memories), but agents still skip it. Make it visually distinct.
5. Run hippo audit — The consolidation pass flagged 7 low-quality memories. A quick hippo audit && hippo prune pass would trim noise and improve recall precision.

Longer-term Themes

• Skill lazy-loading discipline: The most retrieved pattern across dozens of memories is "be lazy with skills — only load when needed." This suggests agents frequently over-load skills preemptively, wasting context budget. Consider a runtime warning when more than N skills are loaded in a single session.
• Validation architecture: Multiple memories reference scratchpad/validation-refactoring.md and the split-at-300-lines rule. This suggests the validation layer is a recurring refactoring target — worth a dedicated cleanup sprint.
• Lock file management: Three verified incidents and the rejig docs #1 CI recall entry all point to .lock.yml management as a systemic pain point. The root cause is that agents edit markdown without running make recompile. A hook-based solution would eliminate this class of errors permanently.
• MCP response size control: Several memories warn about oversized MCP tool responses (code scanning, workflow lists). A systematic audit of all MCP calls in workflow prompts for missing filters would reduce token waste and timeout risk.

Memory Health

• Total memories: 535
• Decayed/removed this cycle: 0 (store is healthy, all memories above decay threshold)
• Low-quality flagged: 7 (run hippo audit for details; candidates for pruning)
• Duplicate semantics deduped: 29 redundant semantic patterns removed this cycle
• Store characterisation: mostly aged, steady operation, mostly episodic, diverse topics

Memory type breakdown

Type	Count
Episodic (observed)	479
Semantic (inferred)	53
Verified incidents	3
Total	535

Consolidation cycle details

Active memories:  535
Removed (decayed): 0
Merged episodic:   79
New semantic:      29
Duplicate semantics deduped: 29
Auto-shared to global: 245
New embeddings:    20 (535/535 total)

References: §27086339636

Generated by 🦛 Daily Hippo Learn · 330.7 AIC · ⌖ 13.6 AIC · ⊞ 20.5K · ◷

• expires on Jun 10, 2026, 7:47 AM UTC

[github-actions[bot]]

This discussion has been marked as outdated by Daily Hippo Learn.

A newer discussion is available at Discussion #38069.