memory-processor

name	memory-processor
description	Iva's daily-memory processor. Reads the day's two-sided transcript (daily/YYYY-MM-DD.md), distills noteworthy entities / decisions / ideas into typed autograph cards, links them into the graph, and produces a daily-summary card (topics + MOC) that navigates down to the raw transcript and up to the week. Model-agnostic — runs on any LLM driving the vault (Iva uses DeepSeek). Triggered by the daily rollup (scripts/memory/rollup.ts daily).
depends_on	["autograph"]

memory-processor — daily transcript → cards + summary

Turn one day of raw conversation into durable, navigable memory.

This skill is judgment-first: the model (you) does the classification, tagging, and linking. The autograph Python scripts are used only for mechanical work (enforce schema, repair links, generate MOCs, decay, touch). enrich.py is never used — you are the enrichment.

Inputs

• daily/YYYY-MM-DD.md — the day's raw two-sided transcript (## HH:MM [text|voice|video|photo|forward from: …] for the user, ## HH:MM [iva] for Iva's replies; older days may use legacy [eva]). See scripts/memory/instructions/rules/daily-format.md.
• schema.json — the vault schema (types, domains, decay).
• Existing cards under cards/** and prior summaries under summaries/, weekly|monthly|yearly/ — for linking and dedup.

Outputs

1. Zero or more entity/decision/idea cards under cards/<kind>/.
2. One daily-summary card at summaries/daily/YYYY-MM-DD.md.
3. A processing marker appended to the raw daily file (never edit existing entries).

Layout & types (from schema.json)

What	Folder	type
Raw transcript (read-only log)	daily/YYYY-MM-DD.md	— (not a card)
Daily summary	summaries/daily/YYYY-MM-DD.md	daily-summary
Weekly / monthly / yearly summary	weekly/, monthly/, yearly/	*-summary
Knowledge note / thought	cards/notes/	note
Person / org	cards/contacts/	contact
Project	cards/projects/	project
Idea / proposal	cards/ideas/	idea
Decision	cards/decisions/	decision

Always pick type and status from schema.json → node_types. Never invent a status.

Flow (4 phases)

1. CAPTURE (phases/capture.md) — read the transcript, segment it, and decide what is noteworthy: which entities, decisions, ideas, and topics the day produced.
2. PROCESS (phases/process.md) — create / update cards for the noteworthy items, choosing exactly one ADD | UPDATE | SUPERSEDE | NOOP operation, then type + description-snippet + tags + status; dedup against existing cards.
3. LINK (phases/link.md) — wire every new card to its domain hub + 2–3 neighbors.
4. SUMMARIZE (phases/summarize.md) — write the daily-summary card: the day's TOPICS plus a MOC linking up to the week, down to the created cards, and down to the raw daily transcript. Then run the mechanical autograph pass.

Mechanical pass (after writing cards & summary)

Run from the project root (Iva's working directory) — the scripts are part of the repo, not of the vault, and take the vault directory as an argument (vault = $ASSISTANT_VAULT_DIR):

# dry-run first, then --apply
uv run scripts/autograph/cleanup.py vault --apply                         # bounded repair before whole-file readers
uv run scripts/autograph/enforce.py vault vault/schema.json --apply   # schema compliance + autofix
uv run scripts/autograph/graph.py fix vault vault/schema.json --apply # repair broken wiki-links
uv run scripts/autograph/engine.py touch vault/summaries/daily/YYYY-MM-DD.md
uv run scripts/autograph/moc.py generate vault vault/schema.json      # regenerate domain MOCs
uv run scripts/autograph/engine.py decay vault                        # recompute relevance/tiers
uv run scripts/autograph/graph.py health vault vault/schema.json      # confirm score

Read .graph/enforce-report.json after enforce.py. Every path in compile_candidates needs semantic repair during the next rollup pass: reread the card, decide its current truth, and use write_card to leave one coherent card. The mechanical pass deliberately refuses to guess when duplicate ## Related sections contain prose.

If uv / Python is unavailable, still produce the cards and summary (they are plain Markdown) and let the nightly Brain run the mechanical pass later.

Hard rules

• Never modify existing transcript entries. Append only a processing marker (see scripts/memory/instructions/rules/daily-format.md).
• No orphans. Every card created here must link to a hub and ≥2 neighbors before you finish (phases/link.md).
• description is a search snippet, not the title. One line, what/why, ~150 chars.
• tags: 2–5, lowercase, kebab-case.
• Idempotent. If the daily file already carries a processing marker and a summaries/daily/YYYY-MM-DD.md exists, only reconcile new entries; do not duplicate cards.
• One structure per card. Exactly one ## Log and one ## Related; never emit dated ## Обновление / ## Update headings. Pass relations only through the write_card.related argument, never inside body.
• Verify writes. Reread every created or updated card before finishing. Confirm one Log, one Related, no empty/dated update headings, and that Compiled Truth says what is true now. A failed invariant keeps the rollup unfinished.
• Quiet days are fine. No noteworthy entities → still write a short daily-summary with topics and the MOC down to the raw transcript. Do not manufacture cards.

References

• references/classification.md — what becomes a card vs. stays in the transcript.
• references/card-templates.md — frontmatter templates per type.
• references/linking.md — hub + neighbor linking protocol.
• references/daily-summary.md — the daily-summary card spec (topics + MOC).
• scripts/autograph/docs/SKILL.md — the typed vault engine (graph, decay, MOC, dedup).
• scripts/memory/instructions/rules/{daily,weekly,monthly,yearly}-format.md — format + rollup chain navigation rules.