feat(memory): learn from tool errors to prevent repetitive failures

[tonbistudio]

Summary

• The ingestion pipeline previously skipped all tool errors (if (tc.error) continue;), so agents never learned from failures ? causing repetitive error loops (e.g., PATH_TRAVERSAL spam for hundreds of turns)
• Now errors are stored as semantic facts under the environment category with keys like tool_error:{tool}:{error_type}
• Repeated errors increment a count and increase confidence (0.5 + 0.1*count, capped at 1.0), making them more prominent in memory retrieval

Details

src/memory/ingestion.ts

• Added normalizeErrorType() helper that maps error strings to consistent labels (PATH_TRAVERSAL, TIMEOUT, PERMISSION_DENIED, RATE_LIMIT, etc.)
• Replaced if (tc.error) continue; with error-learning block that uses existing SemanticMemoryManager.store() UPSERT semantics
• Uses existing environment SemanticCategory ? no type changes needed

src/__tests__/memory.test.ts

• 10 new tests covering: error normalization patterns, semantic fact storage, count incrementing on repeat errors, confidence capping, and error message truncation

Test plan

• npx vitest run src/__tests__/memory.test.ts ? all 83 tests pass (10 new)
• Full test suite ? no regressions (pre-existing failures in command-injection/skills-hardening unrelated)
• Deploy to a running automaton and verify errors appear in semantic_memory table

?? Generated with Claude Code