Ait

Ships the first half of the agent transcript memory pipeline, so what your Claude Code agent said and decided during a session is durable, summarized, and recallable by the next agent — not just the diff.

Added

• Persist Claude Code session transcripts to .ait/transcripts/<attempt-id>.jsonl on session end. They no longer depend on the external ~/.claude/projects/... jsonl staying around.
• Memory policy transcripts block (retain_days, max_total_bytes) controls retention; pruning runs in the daemon reaper. Defaults: 90 days, 500 MB.
• Heuristic transcript summarizer turns each persisted transcript into a compact memory note (topic=transcript-summary) the daemon writes in the background after every attempt_finished.
• transcript-summary:* is added to the default recall_source_allow, so summaries flow through build_relevant_memory_recall into AIT_CONTEXT_FILE automatically.

Migration

Existing repositories carry a frozen recall_source_allow list in .ait/memory-policy.json. To opt into transcript-summary recall, add "transcript-summary:*" to that list, or delete the file and let the next ait init regenerate it.

Design

See docs/agent-transcript-memory-design.md for the full pipeline including not-yet-shipped Steps 3 (non-Claude hook adapters) and the optional LLM summarizer.

Install: pipx install ait-vcs or npm install -g ait-vcs.