Continual learning

Contract

Field	Bound contract
Trigger	A scheduled tick or watcher event fires to mine prior chats and maintain project memory.
Authority	Reversible local: writes only AGENTS.md and the continual-learning index; rollback is version control. No remote mutation.
Side effect	Updates AGENTS.md and the continual-learning index with deduplicated high-signal memory entries.
Done	Deduplicated high-signal memory updates are written, or an explicit no-update result is returned.

Inputs

• Prior chat transcripts or session logs accessible in the local workspace (required).
• Existing AGENTS.md (required, read before mutation).
• The continual-learning index at .continual-learning/index.json (required, read before mutation). The index schema is a JSON object with an array of entries, each carrying fact, source_session, date, and category (one of decision, convention, constraint, resolved-problem, project-knowledge).
• Update scope or focus filter (optional).

Procedure

1. On a scheduled tick or watcher event, enumerate accessible prior chat transcripts and session logs in the local workspace. Done when: every accessible transcript and log is enumerated.
2. Read the current AGENTS.md and .continual-learning/index.json to establish the existing memory baseline. Done when: the existing memory baseline is read and the current set of recorded facts is known.
3. Extract candidate memory facts from the transcripts: decisions, conventions, constraints, resolved problems, and project-specific knowledge. Done when: candidate facts are extracted from every transcript.
4. Deduplicate each candidate against the existing baseline; drop entries that duplicate, contradict without new evidence, or restate lower-signal information already recorded. Done when: every candidate is deduplicated against the baseline.
5. Apply the high-signal gate. A candidate passes when it meets one of: records a decision that changed project direction, establishes a convention or constraint that governs future work, resolves a problem that recurred or is likely to recur, or captures project-specific knowledge not derivable from the codebase. Drop candidates that restate obvious or one-off information. Done when: every surviving candidate is classified and only high-signal entries remain.
6. Capture the prior state of AGENTS.md and the index before writing, so the update can be rolled back. Apply the deduplicated high-signal updates to AGENTS.md and .continual-learning/index.json as local writes only. Done when: the high-signal updates are written and the prior state is captured.
7. If no candidate survives deduplication and the gate, record an explicit no-update result. Done when: a no-update result is recorded or updates are applied.

Failure and recovery

• Unreadable transcript: skip that source, continue with the rest, and report the skipped source in the result.
• Unreadable index: return a blocked result naming the missing or corrupt index; do not write updates without a baseline.
• Conflicting evidence between a candidate and an existing entry: do not overwrite; surface the conflict and leave the existing entry unchanged.
• Partial-result rule: write only the deduplicated subset that resolved cleanly; never write unverified or low-signal entries to meet a quota.
• Rollback: the prior state captured in step 6 restores AGENTS.md and the index to their pre-update content. Revert by replacing the current files with the captured prior state.
• Blocked result: if no transcripts are accessible or the index cannot be read, return a blocked result naming the missing input; do not fabricate memory.

Output

Statement of which deduplicated high-signal memory updates were applied to AGENTS.md and .continual-learning/index.json, or that no update was made and why no candidate survived the gate.