Summary

When an agent fails a task, automatically generate a structured procedural memory entry ("next time, do X when encountering Y") using a separate proposer analysis — the failed agent should not write its own lesson learned.

Motivation

From EvoSkill (arXiv:2603.02766): three-agent separation (Executor/Proposer/Skill-Builder) prevents self-modification bias and produces reusable skills. SynthOrg's MemoryCategory.PROCEDURAL ("Skills & how-to") exists in the design but has no auto-generation mechanism. Current engine/recovery.py classifies failures for reassignment but never extracts learnings.

Design

Failure Analysis Emission

Extend engine/recovery.py to emit a structured failure analysis payload:

• What went wrong (error category, context)
• What capability was missing
• What steps would help next time

Proposer Analysis (Separate from Failed Agent)

A separate LLM call (NOT the failed agent) analyzes the failure and produces a structured procedural memory entry.

SKILL.md Materialization Format

Store as structured entry following Agent Skills format:

• Description + trigger conditions + steps
• Three-tier progressive disclosure (discovery ~100 tokens, activation <5000, execution on-demand)
• Git-native versioning, portable to 30+ agent platforms

Retrieval Integration

Procedural memories surface via memory/retriever.py ranking when agent encounters similar task context in future.

Affected Files

• src/ai_company/engine/recovery.py (failure analysis payload)
• src/ai_company/memory/ (PROCEDURAL category storage + retrieval)
• New: proposer analysis module

Research

• EvoSkill (arXiv:2603.02766) — failure-driven skill discovery
• Agent Skills / SKILL.md (agentskills.io) — materialization format
• SkillNet (arXiv:2603.04448) — skill quality scoring reference

Additional Research (2026-03-26)

Self-Improvement Patterns

XSkill Continual Learning (arXiv:2603.12056):

• Dual-stream architecture: experiences (action-level granularity) + skills (task-level abstraction) extracted from past trajectories
• Reinforces the PROCEDURAL memory category design -- experiences feed skill distillation

Hyperagents Compounding Meta-Improvement (arXiv:2603.19461, Meta AI):

• Self-referential agents where the improvement process itself is editable (DGM-Hyperagent)
• Task agent + meta agent in a single modifiable program
• Autonomously discovers persistent memory and performance tracking
• Results: paper review 0.0->0.710, robotics 0.060->0.372, zero-shot Olympiad math 0.630
• Meta-improvements transfer across domains and compound across runs
• Caution: full autonomous self-modification conflicts with progressive trust; scope to procedural memory/skill layer only

EvoSkill Pareto Selection (already referenced):

• Add detail: Pareto frontier selection retains only non-dominated skill configurations
• +7.3% on OfficeQA, +12.1% on SealQA, +5.3% zero-shot transfer to BrowseComp