Nightly consolidation

Two-stage pipeline: extraction (cheap, local, high volume via Ollama) and consolidation (compresses against a fixed budget). The local model does the dirty work; this skill only orchestrates and interprets.

Why it is designed this way

The typical failure of memory systems is letting a model edit its memory file in place, nightly, on top of its own previous output. Ten days later the file is noise compressed from noise.

Here state/state.json is the source of truth. Aging is deterministic (by date, not model judgment). MEMORY.md is regenerated in full every night. The local model never rewrites its own output: it only produces new entries, compresses an entry to one line when it ages, and proposes links.

Structure

state/
  state.json          # source of truth: entries with date and level
  history/YYYY-MM.md  # cold archive, outside the budget
MEMORY.md             # regenerated artifact — this is what gets read
briefings/            # one concept per day in plain English (local model,
  LATEST.md           #   NOT verified against sources)
knowledge/            # VERIFIED dossiers (optional weekly deep-dive);
  INDEX.md            #   INDEX.md is the catalog
logs/                 # trace of every run

Three levels by entry age: activo (≤7 d, full detail) → contexto (8–28 d, one line) → archivo (>28 d, moved to history/, out of MEMORY.md).

Running the pipeline

Normally the scheduler triggers it. By hand:

py scripts/run.py                 # yesterday (normal nightly use)
py scripts/run.py --date 2026-08-25
py scripts/run.py --days 7        # backfill a week (max ~30: sessions expire)
py scripts/run.py --dry-run       # writes nothing; prints what it would do

Before running, verify Ollama responds: curl -s http://localhost:11434/api/tags. If it doesn't, do NOT improvise with an API model without saying so — local extraction is a privacy decision.

When the user asks for context ("catch me up" / "give me the daily briefing")

Read MEMORY.md in full (it is designed to fit in context) and briefings/LATEST.md; mention the day's concept in one line. If the domain the user is about to work on has a dossier in knowledge/INDEX.md, load it too — verified knowledge beats model memory. Then answer what was asked without dumping the whole file:

• Open items → the abiertos of the active level, grouped by domain.
• "Where did we leave X?" → search X in active and context; if absent, grep state/history/.
• Session start → the active entries of the domain being worked on, plus any cross-links touching it.

Interpreting the run summary

• sesiones: 0 → either no work happened, or claude_projects points wrong. Verify ~/.claude/projects/ has recent .jsonl before assuming the former.
• presupuesto: 4200/4000 → trimming fell short. Fine once; if repeated, raise the budget or lower active_days from 7 to 5.
• json_invalido: N high → model too small for the task; suggest 14B+.
• enlaces: 0 repeatedly → normal in single-domain weeks.

Tuning

Everything configurable is in config.json; prompts live in prompts/ as plain files — edit them there, not in code. When adding a domain, add it BOTH in config.json and in prompts/extract.txt.

An easy mistake to avoid

If the user asks "summarize this week", do NOT re-run the pipeline over days already consolidated — that duplicates entries. Read MEMORY.md; it already is that summary. --days is for backfilling days the scheduler missed.