Mengram 2.47 - Memory that survives /clear, compaction and the next agent

Hi PH, Ali here, solo maker. I launched Mengram seven months ago as "an AI memory API with three types". That was the wrong pitch: nobody buys memory types. What people actually lose is the work. Claude Code compacts and forgets which files it just edited, a new session re-explains the project, Cursor and Codex don't know what Claude decided yesterday, and the next agent redoes a step that was already done. So this version is about that, and it comes with numbers instead of adjectives. What's new since March: - Hooks for Claude Code, Codex and Cursor: recall on every prompt, save after each turn, a checkpoint before compaction that puts the working state back (files you edited, last commands, your last request). - `mengram resume`: a task card written when an agent stops (done, remaining, last test and the commit it ran on), keyed by repo + branch so another session, another agent or another machine picks it up. - Every fact says where it came from (tool, session, date), and a Mac path never reaches a Linux session. - A salience gate before write: 45% less junk stored, no recall lost in 18 runs. - A benchmark you can re-run: over 90 simulated days, 23-62x fewer context tokens than sending the whole history, same or better recall on personal facts. Where it still loses, published: on customer-support dialogue at 30 days recall was 0.25 until this week; the bug was my contradiction pass archiving a loyalty number for a vaguer restatement. It's 0.875 now, and the experiment log with the rejected results is in the repo (experiments/QUEUE.md, RESULTS.jsonl). Install: pip install mengram-ai && mengram setup --key . Or no account at all: the checkpoint and the task card work locally. Apache-2. Question for you: what do you lose most between sessions, the state, the decisions, or the reasons? That's what I'll build next.