Skill Tree Memory Experiments

This repository extends the original MemSkill-style memory workflow with a hierarchical skill tree and markdown-backed negative memories.

The current main experiment target is LoCoMo long-context memory QA. The system learns how to build memory banks from conversation sessions, retrieves memories for test-time QA, and uses negative memories as prompt guardrails for known failure patterns.

What Is Added

• Directory-backed memory skill tree under skills_memory/
• PPO controller routing over hierarchical memory operations
• Markdown negative memory store under negative_memories/
• Automatic training-failure recording as negative memories
• Curated negative-memory clustering and aggregation workflow
• Optional QA-time category matching for all or selected LoCoMo categories
• Hard-case-driven skill-tree evolution during training
• Detailed per-question LoCoMo eval JSON for error analysis
• Standardized experiment scripts under scripts/

Main Scripts

Use the descriptive scripts in scripts/ for new runs:

source ~/.config/skill_tree/env.sh

bash scripts/train_locomo_skilltree_negmem_autoevolve.sh
bash scripts/eval_locomo_skilltree_negmem.sh
bash scripts/sweep_locomo_skilltree_negmem_topk.sh
bash scripts/curate_locomo_skilltree_negmem.sh
bash scripts/eval_locomo_skilltree_negmem_curated.sh

The repository root keeps old script names as compatibility wrappers only.

See:

• RUN_COMMANDS.md for runnable commands and environment setup.
• scripts/README.md for script naming conventions.
• docs/PROJECT_LAYOUT_AND_REMOTE_WORKFLOW.md for repository layout, artifact rules, and local/remote synchronization.
• docs/PROJECT_STATUS_AND_EXPERIMENTS.md for project status, method summary, experiment results, and next steps.

Repository Layout

.
├── data/                    # Small checked-in development datasets
├── docs/                    # Project status, experiment notes, workflow docs
├── negative_memories/        # Template only; generated .md lessons are ignored
├── prompts/                 # Prompt templates
├── scripts/                 # Canonical train/eval/sweep/curation entrypoints
├── skills/                  # Skill-evolution instruction tree
├── skills_memory/           # Memory-operation skill tree
├── src/                     # Core implementation
├── main.py                  # Main experiment driver
├── record_negative_memory.py
├── curate_negative_memories.py
└── RUN_COMMANDS.md

Artifact Policy

Training and evaluation outputs are intentionally not tracked by Git:

• checkpoints/
• logs/
• results/
• wandb/
• curated_negative_memories*/
• generated negative_memories/*.md

These directories are local experiment artifacts. Keep code, scripts, prompts, skill definitions, and documentation in Git; archive large or run-specific outputs separately when needed.