Continuous Learning

Run a file-based learning loop modeled on Clous's memory system, but keep everything inside the repo. Do not rely on embeddings, external databases, or SaaS memory services.

Working Rules

• Keep learning/ at the repo root.
• Prefer updating existing artifacts over creating near-duplicates.
• Capture raw signals first; synthesize later.
• Treat items -> episodes -> triples -> lessons as the promotion ladder.
• Write to identity files and human docs only when the signal is durable enough to outlive the current task.
• Do not turn trivial work into memory noise.

Operating Loop

1. Orient before work

• If learning/ does not exist, initialize it first.
• Read learning/README.md, the most relevant collections, recent lessons, and matching triples before starting.
• Use scan-learning.py to avoid rediscovering known patterns.
• Use references/repo-adaptation.md to choose or refine collection files for the repo type.

2. Capture during work

• Append items as soon as you discover a non-obvious observation, failed attempt, fix, decision, warning, or codebase fact.
• Capture facts while they are still precise. Do not wait until the end of the session.
• Prefer short summaries plus enough detail to make the item reusable later.

3. Consolidate after work

At session end, process signals in this order:

1. Write or update the session episode.
2. Write a decision-trace when the session involved trade-offs, uncertainty, bias checks, or follow-up risks.
3. Extract stable triples.
4. Promote only verified, reusable, non-trivial outcomes into lessons.
5. Update collections, procedures, and beliefs where the knowledge belongs.
6. Refresh learning/README.md and learning/.state/index.json.

4. Promote into source-of-truth docs

After the learning artifacts are written, check whether the knowledge should also change:

• AGENTS.md for stable operating rules and repo facts.
• SOUL.md for persistent collaboration or tone corrections.
• PRINCIPLES.md for decision heuristics.
• Service docs such as README.md, ARCHITECTURE.md, TESTS.md, SETUP.md, or project docs under docs/ when humans should benefit too.

Use references/agents-md-iteration.md for the promotion rules.

Memory Model

Folder Shape

learning/
├── README.md
├── .state/index.json
├── items/
├── episodes/
├── decision-traces/
├── triples/facts.jsonl
├── lessons/
├── collections/
├── procedures/index.md
└── beliefs/current.md

See references/folder-structure.md for schemas and naming rules.

Commands

Initialize the system:

bash .agents/skills/continuous-learning/scripts/init-learning.sh

Capture a signal while working:

python .agents/skills/continuous-learning/scripts/capture-item.py \
  --type discovery \
  --summary "CandidateSerializer owns nested score normalization" \
  --file services/candidates/serializers.py \
  --tag serializer \
  --tag scores

Search memory before starting related work:

python .agents/skills/continuous-learning/scripts/scan-learning.py "serializer normalization"

Refresh the index and incremental state after consolidation:

python .agents/skills/continuous-learning/scripts/refresh-learning.py

Promotion Bar

Promote only when the knowledge is:

• reusable across future sessions
• specific enough to change behavior
• verified by real work, not theory
• non-sensitive
• not already captured more authoritatively elsewhere

Most sessions should update episodes/, triples/, or collections/. Fewer should produce lessons/. Even fewer should change SOUL.md or PRINCIPLES.md.

Anti-Patterns

• Dumping entire transcripts into learning/
• Creating a lesson for every bug fix
• Storing transient branches, commits, or one-off task instructions
• Copying the same fact into items, triples, lessons, AGENTS, and docs without a reason
• Updating identity docs from a single weak signal

Reference Files

• references/folder-structure.md
• references/extraction-patterns.md
• references/agents-md-iteration.md
• references/knowledge-graph.md
• references/memory-types.md
• references/repo-adaptation.md