Thalarch Compound

A completed task may contain knowledge worth keeping. Compounding is the post-task learning gate, not automatic persistence.

1. Candidate lesson

A lesson qualifies only if:

• supported by evidence from this run;
• likely to matter again;
• not obvious from ordinary code reading;
• stable enough to outlive this task;
• compact enough to retrieve without flooding future context.

Examples:

• a hidden build/test prerequisite;
• an ownership/lifecycle invariant;
• a recurring integration contract;
• a reliable diagnostic command;
• a repository-specific convention;
• a failure pattern and its proven discriminator;
• a benchmark/teacher finding that generalizes beyond one case.

2. Distill through experience

For meaningful outcomes, use thalarch-experience to convert the final evidence ledger into a compact card:

trigger → context → problem → discriminator → intervention → evidence → transfer → counterexample.

Preserve useful disproven approaches when they prevent a realistic future mistake. Do not preserve the debugging transcript or private chain-of-thought.

3. Classify memory

Pass the distilled candidate through thalarch-memory:

• IGNORE — weak, obvious, duplicate, noisy, or non-reusable;
• SESSION — useful only to finish/recover the current task;
• PROJECT — durable repository/product knowledge;
• GENERAL — transferable engineering knowledge with a strong generalization basis.

A single successful anecdote should not automatically become GENERAL.

4. Persistence boundary

Store, by default, only in the current work artifact/ledger.

Durable persistence requires an authorized sink. Do not modify AGENTS.md, GEMINI.md, CLAUDE.md, repository docs, .thalarch/brain/, or user-level memory merely because a lesson exists.

When project-local durable memory is explicitly authorized, use thalarch-project-brain and prefer updating/deduplicating an existing entry over appending another near-duplicate.

When a host-native durable memory/RAG system is authorized, use it through thalarch-memory with the same evidence, privacy, scope, freshness, and retirement rules.

5. Privacy and data minimization

Never compound:

• secrets, credentials, auth material;
• private chain-of-thought;
• raw sensitive personal data when a non-sensitive engineering lesson is sufficient;
• entire logs/messages/documents when a compact derived lesson is enough;
• benchmark answer keys or case-specific hacks.

6. Generalization gate

Promote a lesson beyond project scope only when a stable mechanism/contract supports it, multiple independent cases agree, a frozen evaluation/holdout demonstrates broader benefit, or a human explicitly curates it with known limits.

Write durable lessons as:

Context → invariant/lesson → evidence → when to apply → when NOT to apply.

Delete or retire stale lessons. Knowledge bloat is also debt.

7. Teacher/eval handoff

When thalarch-teacher or thalarch-autoresearch produced the evidence, retain both wins and counterexamples. A holdout failure is valuable memory about the boundary of a rule; do not hide it to make the system look smarter.