nicsuzor/remember

Memory Skill

Provides domain instructions for knowledge capture, persistence, and maintenance in the PKB ($ACA_DATA).

Immediate Mode (/remember)

Invariants

• Do not use Write or Edit on files under $ACA_DATA/. Use the PKB MCP toolset.
• Use mcp__pkb__search for semantic discovery. Do not use Glob or Grep on $ACA_DATA/.

Hard Rules

• mcp__pkb__create: Create a document (projects, contexts, meeting notes).
• mcp__pkb__create_memory: Add a memory/note.
• mcp__pkb__append: Extend an existing document.
• mcp__pkb__get_document: Read a document.
• mcp__pkb__search: Search before creating new content — and before asserting a fact. Don't assume; look it up.

Curation

• Capture episodic records accurately in the moment.
• Curate and synthesize knowledge inline as decisions or patterns emerge; do not defer all synthesis to the maintenance loop.

Storage and File Locations

| Content | Tool | Target | | :--------------------- | :---------------------------------------------------- | :---------------------------------- | | Epics/projects | mcp__pkb__create(type="epic"\|"project") | projects/ | | Tasks/issues | gh issue create (GitHub primary) | Synced automatically | | Durable knowledge | mcp__pkb__create or mcp__pkb__create_memory | knowledge/, context/ | | Session findings | mcp__pkb__update_task on parent task | Task body | | Axioms/framework | /framework skill — do NOT write to brain/context/ | skills/ via framework workflow | | Context/goals | mcp__pkb__create | context/, goals/ | | Meeting/call notes | mcp__pkb__create(type="meeting-note") | knowledge/<topic>/ or projects/ | | Sessions/daily | mcp__pkb__create(type="daily-note") | sessions/ |

Episodic classification: Operational records (agent activity, debug logs) → update parent task body only; do not create $ACA_DATA/ docs. Durable episodic (meeting notes, user daily notes) → save to $ACA_DATA/ with correct type frontmatter.

Canonical Topic Notes (Enduring Memory)

Maintain exactly one canonical note per first-class topic (tools, projects, skills, agents, concepts).

• Update existing notes with mcp__pkb__append. Create new notes only when the topic does not exist.
• Avoid one-off observation files.

Reconciliation

Search for duplicate or peer notes during updates. Merge unique content, retire weaker versions, update wikilinks. Delete superseded files — do not leave superseded_by: pointers; follow [[consolidation-procedure]] Step 3.

Workflow

1. Search: mcp__pkb__search(query="topic").
2. Evaluate: If a canonical note exists, append to it. Do not create a separate file.
3. Write: Call mcp__pkb__create(title, body, type, tags) or mcp__pkb__create_memory(title, body, tags).

Graph Integration

• Every note must contain at least one [[wikilink]] to a parent concept, project, goal, or entity in prose.
• Link all proper nouns; provide markdown links for external issues and PRs (e.g., [org/repo#123](url)).
• List semantic relationships in a ## Relationships section: - [related] [[task-id]] — description.
• Wikilinks: prose only — never inside code blocks, inline code, or technical tables. No standalone "See Also" section.
• Abstraction: Capture generalizable patterns, not local implementation details.
• Observation callouts (required for factual claims in episodic notes):

  > [!observation] Factual statement
  > Source: [[source-note]]
  > Confidence: <0.0–1.0>  (≥0.8 established · 0.4–0.79 provisional · <0.4 speculative)
  

  If a new observation contradicts an existing one, record both with sources and flag. Do not overwrite silently.
• type: knowledge provenance (required): sources:, synthesized:, last_reviewed:, confidence:, maturity: seedling|budding|evergreen.
• MOCs: Create when a topic has 5+ notes; use type: moc, tags: [moc, topic-area], created:, last_reviewed:.

Background Capture

To capture asynchronously, spawn a model="haiku" background agent with run_in_background=true and prompt it to invoke Skill(skill='remember') with the content.

Output Expectations

State PKB write details directly: tool used, note title, returned ID. Do not expose raw filesystem paths.

Maintenance Mode (/sleep)

The canonical home of the former /sleep skill. Runs periodic offline consolidation on a GHA workflow or manual trigger.

Triggers

Invoke Maintenance Mode on: explicit /sleep call; GHA schedule or workflow_dispatch; or SLEEP_MODE env var set.

Rules

• Never fabricate assertions without provenance.
• Do not edit daily notes, meeting notes, or task bodies (except adding consolidated: YYYY-MM-DD in frontmatter).
• Record and preserve contradictions.

Mode Detection

First match determines mode:

1. SLEEP_MODE env var overrides (short-loop | full-session).
2. LOOP_INTERVAL_MINUTES <= 30 → short-loop.
3. Prior cycles ran in this session → short-loop.
4. Default → full-session.

Maintenance Phases

See [[references/maintenance-phases]] for sub-agent rules, pacing, and exit protocols.

| Phase | Name | | :---- | :-------------------------- | | 0 | Graph Health Baseline | | 1 | Session Backfill | | 2 | Transcript Mining | | 3 | Episode Replay | | 4 | Knowledge Consolidation | | 5 | Index Refresh | | 6 | Data Quality Reconciliation | | 7 | Staleness Sweep | | 8 | Refile Processing | | 9 | Graph Maintenance | | 10 | Consolidation Self-Check | | 11 | Brain Sync |

CI Environment Constraints

On GitHub Actions, no PKB MCP is available. Use direct file operations (Bash, Glob, Grep, Read, Write, Edit) on markdown files. Do not edit files outside $AOPS_SESSIONS/ (except git commits/push).