Answer
One thing. Honcho is the memory service behind my recall. Today's work had one goal: make it store durable facts about you instead of my own play-by-play. One decision is still open: which stored copy of the Honcho key is authoritative. Either the block in honcho.json, or the copy 1Password injects into the environment.
- My seat wrote 40,021 conclusions today; 28,619 of them (71.5%) were me describing my own turns, not facts about you.
- The fix for that is a workspace-level extraction instruction you applied at 08:55. It cannot be set per seat, which is why it was a fleet decision.
- The dream job exists only because Honcho will not consolidate while you keep talking; it needs a local tick to ask.
- Deepseek's review found the tick can spend without permission in two ways. That is a ten-line fix, not a redesign.
Applies toJames seatworkspace franck-m42026-11-07
Where your memory actually lives
Every turn passes the same four pieces: a key, a config file, the plugin, and the hosted workspace. Nothing else in this thread touches your memory.
Diagram structure
- 1Password vault
vault - Process environment
env - honcho.json block
json - Honcho plugin
plugin - Workspace franck-m4
ws - Dream cron tick
tick vault→env· read once, cached 5 minenv→plugin· key (current winner)json→plugin· key and settingsplugin→ws· save and recalltick→json· reads keytick→ws· schedule_dream
The plugin prefers the key in honcho.json; my tick script prefers the environment. Both work, but they can disagree after a re-key — that is the whole open question.
What actually changed today
| Change | Where | Effect | Status |
|---|---|---|---|
| Latency and budget tuning | honcho.json, my block only | turn → async writes, timeout 120 → 60 s, context 1,200 → 3,000 tokens | Done and verified |
| Extraction instruction | workspace franck-m4, all 38 peers | deriver told to keep durable facts only | Applied at 08:55; efficacy unmeasured |
| Peer card limit | same workspace payload | roughly 12 entries, drop stale ones | Applied with the above |
| SDK version | shared runtime | 2.5.1 → 2.8.0, upgraded by Deepseek | Verified; 2.9.0 not yet published |
| Peer audit and export | read-only | 38 peers, 1,150 sessions, 240,824 conclusions inventoried | Preserved at /Volumes/T7/archives/honcho-audit-20261107/ |
- Conclusions my seat wrote
- 40,021
- Of those, about myself
- 28,619 (71.5%)
- About you
- 11,390 (28.5%)
- Dream tick fires since September
- 617 — activity, not proof of dreams
The one decision
| Criterion | honcho.json wins | 1Password wins |
|---|---|---|
| A fire needs no vault lookup | independentYes | No, but cached |
| Survives a re-key automatically | manual editNo | automaticYes |
| Matches the plugin's own precedence | alignedYes | divergesNo |
| Extra 1Password reads per fire | Zero | Zero — the parent already resolved them |
48reads/h / 1000reads/h
Measured shape, not a counter: 4 secrets per 5-minute cache window. Business tier allows 10,000/h; Families and Teams 1,000/h, which is the limit shown.