Grounded in the message
Every proposal cites the source message it came from. No claim makes it to memory without a pointer back to where the session actually said it.
A coding session is where the learning happens — and where it's usually lost. Memophant imports your coding-agent transcripts into the repo and distills them into durable notes you approve one at a time. The session ends; the insight stays, versioned and queryable.
Distillation runs on Claude — it uses the Anthropic API, so it needs your own key (kept in your macOS Keychain). Importing and reading your sessions works without it.
How it runs
Point Memophant at a session and it works the transcript in stages: configure the run, watch it scan, review every proposed note as a before-and-after, and apply only what's right. Each step is reviewable; nothing is written silently.

The guardrails
Batch 'ingest everything' passes produce low-signal notes and discourage upkeep. Distillation is the opposite by design: few candidates, every one cited, the spend bounded.
Every proposal cites the source message it came from. No claim makes it to memory without a pointer back to where the session actually said it.
A distill pass proposes only a handful of candidates per session, not a noisy ingest of everything. The lesson survives; the chatter doesn't.
A configurable token budget caps how much of a long transcript gets sent, so the API spend never runs away on a marathon session.

The full transcript, too
Distillation pulls the durable lessons, but the source transcript is kept as well — converted to readable markdown with collapsible tool calls and per-message anchors, so any citation can be traced back to the exact moment it was said.
Every long session can leave behind a durable note instead of a dead transcript.