🦛 Hippo Memory Insights — 2026-06-12

[github-actions[bot]]

Summary

• Memories in store: 535 (482 episodic, 53 semantic)
• New lessons learned from commits today: 0 (no fix/revert/bug commits in the last 7 days)
• New semantic memories created: 29 (merged from 79 episodic clusters)
• Embeddings refreshed: 17 new, 535/535 total
• Highest-confidence memory: mem_4afaa0965161 — Stale .lock.yml files cause CI friction; run make recompile after markdown workflow edits (verified incident, 36x retrieved, strength 1.00)

Top Memories Surfaced

1. [62x] mem_cab727bbe2e5 — gh-aw identity: gh aw compiles markdown workflows into GitHub Actions; it is NOT the GitHub Copilot CLI. Keep these separate in documentation and error messages.
2. [62x] mem_0324334270d0 — MCP safety: Always guard list_code_scanning_alerts calls with state: open and severity: critical,high to avoid oversized MCP responses.
3. [62x] mem_3c0154ff3f2d — Post-merge discipline: Run make build && make recompile after merging main to rebuild the binary and recompile all workflow lock files.
4. [52x] mem_2704ae961669 — Checkpoint 1: make build && make fmt immediately after the first substantial code change — surfaces compile errors before spending more context on subsequent edits.
5. [36x] mem_4afaa0965161 (incident, verified) — Stale .lock.yml files are the rejig docs #1 recurring CI friction source; make recompile must follow every markdown workflow edit.
6. [19x] mem_e02072c9ae07 (incident, verified) — node: command not found on GPU/self-hosted runner paths; validate runtimes and PATH early.
7. [50x] mem_512f4cbb0103 — ANSI stripping: The workflow compiler (pkg/workflow/compiler_yaml.go) automatically strips ANSI codes via stringutil.StripANSI() from descriptions, sources, and comments — no manual stripping needed.

Suggested Improvements

Error Patterns

• Stale .lock.yml files (recurring, verified): This is the most-retrieved incident. Consider adding a pre-commit hook or CI gate that verifies lock files are up-to-date after any .md workflow change. The memory has been retrieved 36 times, signalling repeated exposure.
• MCP response overflow: list_code_scanning_alerts without filters consistently causes oversized responses. Add a linter rule or prompt guard to always inject state: open + severity: critical,high.
• Unformatted Go code causing CI failures: Five CI failures in a single day were attributed to skipping make fmt. Enforce Checkpoint 1 validation in agent prompts.
• Codex auth failures (mem_b3de87b76481, decaying): Verify auth/mode/token setup before triggering reruns on GPU/self-hosted environments. This memory is nearly dead (strength 0.01) — the issue may be resolved or forgotten.

Code Quality

• File splitting discipline: The validation-refactoring.md scratchpad and associated memories show a pattern of oversized files. Files >300 lines or spanning 2+ distinct domains should be split. The skills/javascript-refactoring/SKILL.md guide is the canonical reference.
• Type safety patterns: Using distinct named types (e.g., JobName vs StepID) prevents accidental mixing — 50x retrieved, suggesting agents frequently benefit from this reminder. Consider codifying more of these patterns in AGENTS.md.
• Prefer new files for new features over extending existing ones — a structural principle that surfaces frequently (22x via code-organization tag).

CI Health

• Checkpoint 2 ("before every PR") is the most-tagged category (65 memories) — its high recall frequency suggests agents are checking it repeatedly. Consider consolidating those memories further or pinning the most-actionable ones.
• make test is slow (>5 min): Agents should rely on CI for the full test suite. Only make test-unit (~3 min) is appropriate during development.
• validate-yaml ANSI scan runs as a pre-gate in CI — any ANSI leak in YAML files will block all other jobs. The compiler strips these automatically for workflow files, but hand-edited YAML still needs care.

Quick Wins

1. Add make recompile to a post-merge checklist or Makefile alias — resolves the rejig docs #1 CI friction source (stale .lock.yml) with near-zero effort.
2. Pin mem_0324334270d0 (guard list_code_scanning_alerts) — it is retrieved 62x but unpinned; pinning prevents accidental decay.
3. Prune mem_b3de87b76481 (Codex auth failures, strength 0.01) — this memory is nearly dead and never retrieved since creation; safe to remove.
4. Add a workflow prompt note reinforcing that make build && make fmt (Checkpoint 1) must be run before any additional edits — 52x retrieved but not yet a pinned rule.
5. Run hippo audit to address the 7 low-quality memories flagged during consolidation.

Longer-term Themes

• Episodic dominance (482/535 = 90%): The store is heavily episodic. The 79 merges during this sleep cycle created 29 new semantic memories, but 240 low-strength memories remain. A focused pruning/promotion pass could tighten the store significantly.
• Checkpoint 2 fragmentation: 65 memories share the checkpoint-2 tag — many are near-duplicates of "run make agent-report-progress and verify zero lint errors before every PR." A targeted consolidation pass on this cluster would reduce recall noise.
• Skill lazy-loading: The important-using-skills tag (12 memories) consistently surfaces the BE LAZY principle. Consider elevating this to a pinned rule to reduce redundant skill-loading in agent sessions.

Memory Health

Metric	Value
Total memories	535
Episodic	482
Semantic (inferred)	53
Low-strength (< 0.4)	240
Incidents (verified)	3
Low-quality flagged	7
Shared to global store	245

Low-strength incidents worth pruning

• mem_b3de87b76481 — Codex auth failures (strength 0.01, never retrieved since creation 23 days ago). Safe to delete.

Consolidation details from this run

• Merged 79 episodic entries → 29 new semantic patterns
• Deduped 29 redundant semantic patterns (kept stronger copies)
• 7 low-quality memories flagged (run hippo audit to review)
• 245 high-value memories auto-shared to global store
• 17 new embeddings (535/535 total)

References:

• Workflow run: §27402612367

Generated by 🦛 Daily Hippo Learn · 321.2 AIC · ⌖ 33 AIC · ⊞ 20.5K · ◷

• expires on Jun 15, 2026, 12:03 AM UTC-08:00