memory: track agent memory in repo .claude/memory (symlinked from harness path)

Move per-project agent memory into the repo so it is version-controlled.
Harness path /root/.claude/projects/.../memory now symlinks here, so
auto-load still works while the files live under git. Un-ignore only
.claude/memory/ (worktrees, locks, local settings stay ignored).

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

.claude/memory/MEMORY.md

@@ -0,0 +1,4 @@
+- [AFK autonomy](feedback_afk_autonomy.md) — during AFK, prefer queueing follow-ups over standing down; reserve "stop and ask" for craft-heavy moves.
+- [No nohup with pueue](feedback_no_nohup_with_pueue.md) — run `pueue follow|wait` directly as the bg task; nohup&  orphans it from the harness.
+- [Burn down task list](feedback_burn_down_task_list.md) — when many asks are queued, do them all; don't stop to ask which first.
+- [Workshop paper goal](project_workshop_paper_goal.md) — current phase is ablations+seeds for a workshop paper; artifact tracker A1-A7 lives in docs/spec/20260602_writeup_spec.md.

.claude/memory/feedback_afk_autonomy.md

@@ -0,0 +1,14 @@
+---
+name: feedback-afk-autonomy
+description: "When the user is AFK (via /afk command or AFK cron pings), prefer more autonomous queueing not less. Do not stand down on speculative work just because the headline question is answered."
+metadata: 
+  node_type: memory
+  type: feedback
+  originSessionId: c06c2b0e-44ee-45a1-b213-4f77c006109c
+---
+
+When the user is offline (AFK), the default should be to fill the queue with reasonable follow-ups, not to stand down and wait for explicit direction. The user prefers autonomous initiative during AFK windows.
+
+**Why:** The user said, after I had explicitly chosen to stop queueing speculative work mid-AFK: "actually I'd prefer the opposite if I'm afk". The CLAUDE.md preference for "stop and ask" applies during interactive sessions; during AFK windows, time is the scarcer resource than judgment calls.
+
+**How to apply:** During /afk-driven or cron AFK pings, when the immediate question is answered and the queue is empty, do NOT default to "standing down until you're back". Instead, queue the natural next batch (additional seeds for confidence, an ablation that resolves the obvious follow-up question, a sweep over a parameter that's now informed by the result). Pick the option with the best information-per-dollar ratio that doesn't require non-trivial code edits or new craft work. Reserve "stop and ask" for moves that genuinely need user judgment (e.g. drafting hand-crafted training data, rewriting non-trivial code, changing experimental direction). See [[afk-protocol]] in this memory store if it exists.

.claude/memory/feedback_burn_down_task_list.md

@@ -0,0 +1,14 @@
+---
+name: feedback_burn_down_task_list
+description: "When several asks/tasks are queued, do them all; don't stop to ask which to do first."
+metadata: 
+  node_type: memory
+  type: feedback
+  originSessionId: c06c2b0e-44ee-45a1-b213-4f77c006109c
+---
+
+When the user has given multiple asks or there's a task list, work through all of them rather than stopping to ask "which should I do next / shall I do X?". Asking-to-pick when work is already queued is the worst move: it does zero work AND blocks communication.
+
+**Why:** the user often goes AFK and trusts autonomy; a "which first?" question stalls everything for nothing. They've said this more than once (see [[feedback_afk_autonomy]]).
+
+**How to apply:** address/brainstorm/fix every open point in one pass; only stop to ask when there's a genuine fork with no defensible default or an irreversible/outward-facing action. "Want me to proceed?" on already-requested work is not such a fork.

.claude/memory/feedback_no_nohup_with_pueue.md

@@ -0,0 +1,23 @@
+---
+name: feedback-no-nohup-with-pueue
+description: "Don't wrap pueue follow/wait in `nohup ... &`; run the client directly as the background task."
+metadata: 
+  node_type: memory
+  type: feedback
+  originSessionId: c06c2b0e-44ee-45a1-b213-4f77c006109c
+---
+
+When monitoring pueue jobs from a `run_in_background` bash task, run the client
+directly: `pueue follow <id> | tail -N` (or `pueue wait <id>; pueue log <id>`).
+Do NOT wrap it in `nohup bash -c '... &'`.
+
+**Why:** pueue's daemon already owns the job; `follow`/`wait` are just clients
+streaming the daemon's log, so nohup buys nothing. Inside `run_in_background`
+(itself already a harness-tracked process), an extra `nohup ... &` spawns a
+child and the outer shell exits in ~0s, so the harness notifies on the wrapper
+exit while the real follow runs orphaned. Symptom: "completion notifications
+fire early / abort my sleeps."
+
+**How to apply:** background command = the bare pueue client, no nohup, no `&`.
+The harness then tracks the client itself and wakes me on real finish/fail.
+Related: [[feedback-afk-autonomy]].

.claude/memory/project_workshop_paper_goal.md

@@ -0,0 +1,24 @@
+---
+name: project-workshop-paper-goal
+description: Current phase of projected_grpo -- filling out ablations/seeds for a workshop paper; the canonical artifact tracker and which runs feed each.
+metadata: 
+  node_type: memory
+  type: project
+  originSessionId: c06c2b0e-44ee-45a1-b213-4f77c006109c
+---
+
+As of 2026-06-02 the method (route2 gradient routing vs reward hacking) works at
+n=1 and the project is in "fill out ablations + seeds for a workshop paper/blog"
+mode. The canonical, durable goal tracker is docs/spec/20260602_writeup_spec.md
+(force-added to git despite /docs/ being gitignored). It lists artifacts A1-A7:
+
+- A1+A2 keynote fig + table: route2 vs vanilla n=3 deploy hack/solve (task #173). Blocked on vanilla band jobs 74+79.
+- A3 ablation table: vanilla / erase / route2-full / frozen-vs-refresh / random-V(81) / placebo(80) / post-hoc tt-erase(83) (task #183).
+- A4 long-run fig: 200-step route2(77) vs vanilla(82) (task #184).
+- A5 generalisation (the no-cheat payload C2): held-out-mode suppression, NOT YET QUEUED, highest-value gap (task #185).
+- A6 appendix traces (4 loophole classes): done, blog appendix.
+- A7 appendix ablation context: results.md Q-rows already exist.
+
+Backup = out/runs (RL results, not regenerable) + out/pools + out/pairsets +
+out/vhack + the curated docs/. Skip out/baked (23G) and out/vhack_grads (1.8G):
+deterministic caches that rebuild. See [[project-out-dir-backup]] if written.

.gitignore

@@ -1,4 +1,5 @@
-.claude/
+.claude/*
+!.claude/memory/
 .venv/
 /out/
 /data/