remember

Distill durable lessons into .ultragoal/memory/. Focus given by the user (may be empty — then review the whole session and any just-finished goal, especially its Decision journal and Verification log):

$ARGUMENTS

The two layers — never mix them

Every memory file has two layers separated by ---:

• Above the line — compiled truth. The current understanding, rewritten freely as it improves. This is what future sessions read first.
• Below the line — evidence log. Append-only, dated, never rewritten or deleted. One self-contained line per observation: [YYYY-MM-DD S<n>] <what was observed and how>.

Synthesis can be wrong; evidence can't. When you update a compiled claim, the evidence that justified the old version stays in the log. Anyone can re-derive the truth from below the line — that's the point.

Where things go

Each file's header says what belongs in it and what doesn't — read it before filing. The short version:

• facts.md — what is true of this repo (schemas, behaviors, invariants, tool quirks)
• patterns.md — reusable approaches that worked here, with why they worked
• failures.md — dead ends: what was tried, why it failed, what to do instead
• MEMORY.md — the index and the fixed slots (commands, invariants, gotchas, hot files). Fill [no data yet] slots as you learn; keep one index line per entry. Its head is injected into every session — keep it tight.

Unsure where something goes? File the evidence line in the closest file and flag it in the index rather than inventing a new home. The same fact living in two places is how agents stop trusting the system.

Provenance — every compiled claim carries how it's known

[VERIFIED S4 · ran: pnpm test · 2026-06-10]   you (or the verifier) executed a command and saw it
[READ S2 · docs/auth.md]                      a source said it; the source could be stale
[INFERRED S5 · medium]                        you concluded it; it has never been directly checked
[USER-CORRECTION · 2026-06-10]                the user told you; highest confidence there is

Rules:

• Never upgrade a tag without new evidence, and never let synthesis launder INFERRED into unqualified fact — confident prose is indistinguishable from verified truth once written down, so the tag is the only thing keeping the two apart.
• A wrong [VERIFIED] claim is worse than no claim: if evidence contradicts one, fix it above the line and append the contradicting evidence below.
• User corrections are written immediately, the moment they happen — not batched to the end of the goal. They override everything, they're the highest-confidence signal you will ever receive, and they die with the session if you defer them.

What to save

• One lesson per entry, one-line summary first; record corrections and confirmed approaches, including why they mattered.
• Expensive derivations you'd hate to redo — a traced data flow, a debugged race, a mapped dependency — file them the moment you finish deriving them; they shouldn't evaporate into chat history.
• Don't save what the repo or git history already records. Save the non-obvious: what you'd want the next session to know on minute one.
• Generalize when evidence supports it: three observations sharing a cause become one compiled rule, with the evidence lines still below the line, each dated.

When done, list for the user what was added, updated, or deleted — one line each.