mirror-v0

structure (daily and weekly)

daily (runs at midnight via cron)

purely observational. no llm involved.

collects:

• notes created (new .md files in vault)
• notes edited (modified timestamps)
• sessions that occurred (from session data)
• orphan notes flagged (notes with no backlinks, detected that day)

writes to: mirror-outputs/daily/YYYY-MM-DD.md (outside vault, in research-infrastructure repo or dedicated folder)

if no activity: writes a minimal dated file recording the absence. absence is data.

weekly (runs sunday midnight via cron)

draws on: all seven daily files for the window + a fresh chromadb pass over vault activity for the week.

structured operations (no llm):

• frequency/recurrence counts: tags, links, phrases appearing most across the week’s notes
• notes created/edited this week, lightly grouped by folder/topic
• orphans accumulating across the week

llm synthesis (local model – qwen2.5:32b recommended, llama3.3:70b as fallback):

two passes, both shown with their grounding data visible in the output so the ai’s reasoning is transparent and auditable.

pass 1 – synthesis statement:

• input: top recurring terms, orphan list, notes created/edited this week (titles and folders)
• output: a short paragraph describing what the week’s notes seem to be circling around
• the grounding data is displayed alongside the statement in the output (not hidden)

pass 2 – single generated reflective prompt:

• input: the synthesis statement just produced + top recurring terms + orphan list
• instruction to model: “generate one specific question grounded in the above. name actual notes, terms, or patterns you can see in the data. do not ask generic reflective questions.”
• output: one question only
• specificity constraint does significant work regardless of model quality – the model follows the instruction rather than needing to be clever
• if the prompt misses it costs little; one question is easier to ignore than three and less damaging when it fails

on model quality for this task: the bar is not a brilliant insight – it is a prompt that causes a momentary pause rather than being skipped. that is achievable with qwen2.5:32b given well-prepared, constrained input. if generated prompts are consistently poor after testing, this moves to v1 and is replaced with no prompt (synthesis statement only) in v0.

writes to: mirror-outputs/weekly/YYYY-WNN.md

output format

dated markdown files in mirror-outputs/ folder. outside the vault. not indexed by obsidian. readable by you; you decide what if anything comes back into the vault. the act of curation is yours.

folder structure:

mirror-outputs/
  daily/
    2026-06-23.md
    2026-06-24.md
    ...
  weekly/
    2026-W26.md
    2026-W27.md
    ...

interface integration

phd-live

mirror surfaces via a toggle on the existing live column (not a separate page). toggling switches the live column into mirror mode.

mirror mode shows:

• most recent daily digest
• most recent weekly digest
• history of past digests (scroll)

learning dashboard (react/vite)

mirror gets its own tab in the existing dashboard.

dashboard tab shows:

• daily digest feed
• weekly digest with pattern outputs and reflective prompts visible

cron jobs

# daily mirror run -- midnight
0 0 * * * /path/to/mirror/run_daily.sh

# weekly mirror run -- sunday midnight
0 0 * * 0 /path/to/mirror/run_weekly.sh

explicitly out of scope for v0

• bot activity logging (next todo after v0)
• episodic/a-mem style memory (v1 consideration – simple append log approach most likely)
• frontier model access
• vault write access (permanent constraint, not a v0 limitation)
• fine-tuning local models for mirror tasks

local model notes

qwen2.5:32b: recommended for weekly synthesis – good balance of capability and speed on 32gb ram. llama3.3:70b-instruct-q4_K_M: stronger reasoning, significantly slower – use if weekly run timing allows. the quality gap relative to frontier models is documented and framed as a research finding (consistent with study companion precedent).