Remember Skill

Taxonomy note: This skill provides domain expertise (HOW) for knowledge capture and persistence. See [[TAXONOMY.md]] for the skill/workflow distinction.

Persist knowledge to markdown + PKB. Both writes required for semantic search.

Memory Model

$ACA_DATA contains both semantic and episodic memory. The key distinction is between synthesized knowledge (decontextualized, kept current) and primary sources (time-stamped, preserved as-is).

Semantic Memory (synthesized knowledge)

Durable, decontextualized truths. Lives in $ACA_DATA/knowledge/, project files, context files.

• What IS true now. Understandable without history.
• If you must read multiple files or piece together history to understand truth, it's not properly synthesized.
• Always cites its episodic sources (see Provenance below).

Episodic Memory (three types, all legitimate in $ACA_DATA)

1. Task bodies (type: task): Document what was done. Preserved even when archived. Managed via tasks MCP.
2. Daily notes (type: daily-note, in sessions/): High-quality user synthesis of what happened and what matters. Created by the user. NOT edited after the day.
3. Contemporaneous notes (type: meeting-note, in knowledge/ or project dirs): Notes of meetings, phone calls, conversations. Captured close to the event. May not be edited afterwards. Valuable as primary sources.

Cognitive Foundations

The episodic/semantic distinction mirrors how biological memory works. Complementary Learning Systems theory (McClelland et al. 1995) shows that rapid episode capture and gradual pattern extraction are complementary processes — you need both. Semanticization (Baddeley 1988) is the natural process where episodic memories lose temporal context through repeated retrieval, becoming context-free semantic knowledge. The /sleep cycle's consolidation phases mirror this: offline replay of episodes, pattern extraction, integration into durable knowledge. The review process IS the consolidation mechanism — passive storage does not produce understanding.

Synthesis Flow (the consolidation pipeline)

• Episodic content accumulates → patterns emerge across multiple notes
• Agent or human extracts observations → creates/updates semantic knowledge notes
• Semantic notes always cite their episodic sources (provenance)
• The transformation from episodic to semantic mirrors cognitive semanticization
• Git history preserves the full record; semantic notes in $ACA_DATA reflect what's current

Storage Hierarchy (Critical)

PKB is the universal index. Write to your primary storage AND PKB for semantic search retrieval.

What	Primary Storage	Also Sync To
Epics/projects	PKB (type="epic" or type="project")	PKB index
Tasks/issues	GitHub Issues (gh issue create)	PKB index
Durable knowledge	$ACA_DATA/ markdown files	PKB index
Session findings	Task body updates	PKB index

See [[base-memory-capture]] workflow for when and how to invoke this skill.

Decision Tree

Is this a time-stamped observation? (what agent did, found, tried)
  → YES: Use tasks MCP (create_task or update_task) - NOT this skill
  → NO: Continue...

Is this about the framework (axioms, heuristics)?
  → YES: HALT and invoke /framework skill to add properly to $AOPS
  → NO: Continue...

Is this about the user? (projects, goals, context, tasks)
  → YES: Use appropriate location below
  → NO: Use `knowledge/<topic>/` for general facts

File Locations

Content	Location	Notes
Project metadata	projects/<name>.md	Hub file
Project details	projects/<name>/	Subdirectory
Goals	goals/	Strategic objectives
Context (about user)	context/	Preferences, history
Sessions/daily	sessions/	Daily notes only, type: daily-note
Tasks	Delegate to [[tasks]]	Use scripts
General knowledge	knowledge/<topic>/	Facts NOT about user
Meeting/call notes	knowledge/<topic>/ or projects/	Contemporaneous notes, type: meeting-note
Maps of Content (MOCs)	knowledge/ or topic dirs	Navigational hub notes, type: moc

Episodic Content → Where It Belongs

Use Tasks MCP (NOT $ACA_DATA files)

• Individual agent actions: "Completed X on DATE" → mcp__pkb__create_task(task_title="...", type="task", project="<project>", parent="<parent-id>")
• Debugging logs: "Discovered bug in Y" → mcp__pkb__create_task(task_title="...", type="task", project="<project>", parent="<parent-id>", tags=["bug"])
• Experiment step-by-step records: "Tried approach A" → mcp__pkb__update_task(id="...", body="...")

Rule: If it describes agent activity or debugging, it's operational episodic → tasks MCP.

Episodic Content in $ACA_DATA

• Daily notes (user-created summaries in sessions/) — type: daily-note
• Meeting/call notes (type: meeting-note) — contemporaneous records of conversations, captured close to the event
• Contemporaneous observations that may not be edited later — primary sources valued for their accuracy at the time of capture

Workflow

1. Search first: mcp__pkb__search(query="topic") + Glob under $ACA_DATA/
2. If match: Augment existing file
3. If no match: Create new file with frontmatter:

---
title: Descriptive Title
type: note|project|knowledge|moc|meeting-note|daily-note
tags: [relevant, tags]
created: YYYY-MM-DD
---

Content with [[wikilinks]] to related concepts.

4. Sync to PKB:

mcp__pkb__create_memory(
  title="[descriptive title]",
  body="[content]",
  tags=["relevant", "tags"]
)

Graph Integration

• Every file MUST [[wikilink]] to at least one related concept
• Project files link to [[goals]] they serve
• Knowledge files link proper nouns: [[Google]], [[Eugene Volokh]]
• Semantic Link Density: Files about same topic/project/event MUST link to each other in prose. Project hubs link to key content files.

External References (REQUIRED)

When a memory references an external issue, bug, or resource, always link it explicitly:

• Upstream bugs: [org/repo#NNN](https://github.com/org/repo/issues/NNN) — don't just mention "#NNN" in prose
• Internal issues: gh issue create link or [#NNN](url)
• Related PKB nodes: Add a ## Relationships section with typed edges:

  ## Relationships
  - [related] [[task-id]] — brief description
  - [upstream-bug] [org/repo#NNN](url)
  - [parent] [[parent-id]]
  

Why: Unlinked references are dead ends. The PKB graph and future agents can't traverse prose mentions — they need explicit edges.

Wikilink Conventions

• Wikilinks in Prose Only: Only add [[wikilinks]] in prose text. Never inside code fences, inline code, or table cells with technical content.
• Semantic Wikilinks Only: Use [[wikilinks]] only for semantic references in prose. NO "See Also" or cross-reference sections.

Semantic Search

Use PKB semantic search for $ACA_DATA/ content. Never grep for markdown in the knowledge base. Give agents enough context to make decisions - never use algorithmic matching (fuzzy, keyword, regex).

Abstraction Level (CRITICAL for Framework Work)

When capturing learnings from debugging/development sessions, prefer generalizable patterns over implementation specifics.

❌ Too Specific	✅ Generalizable
"AOPS_SESSION_STATE_DIR env var set at SessionStart in router.py:350"	"Configuration should be set once at initialization, no fallbacks"
"Fixed bug in session_paths.py on 2026-01-28"	"Single source of truth prevents cascading ambiguity"
"Gemini uses ~/.gemini/tmp/<hash>/ for state"	"Derive paths from authoritative input, don't hardcode locations"

Why this matters: Specific implementation details are only useful for one code path. Generalizable patterns apply across all future framework work. We're dogfooding - capture what helps NEXT session, not what happened THIS session.

Test: Would this memory help an agent working on a DIFFERENT component? If not, it's too specific.

Observation Notation

When extracting facts or observations from episodic content, use Obsidian callout syntax:

[!observation] Brief factual claim Source: [[link-to-source-note]] or description of origin Confidence: established | provisional | speculative

Examples:

[!observation] Platform liability frameworks increasingly distinguish between hosting and curation Source: [[20260401-meeting-regulators]], discussion with policy team Confidence: established

<!-- -->

[!observation] Sleep cycle deduplication catches ~15% false positives on short titles Source: Three /sleep runs in March 2026 Confidence: provisional

Guidelines:

• One fact per observation block
• Always include source — never assert facts without provenance
• Confidence levels: established (multiple independent sources), provisional (single source or limited evidence), speculative (inference, needs verification)
• Observations in episodic notes (daily, meeting) are raw material; observations in knowledge notes are synthesized claims
• Humans may also write observations informally as plain prose — the callout format is a recommendation, not a requirement
• Contradictions: When a new observation contradicts an existing one, record BOTH with their sources. Never silently overwrite. Flag for human resolution. This prevents catastrophic forgetting — schema-inconsistent information must be integrated gradually, not by replacement.

Provenance

All synthesized knowledge must be traceable to its sources. This is critical — we never fabricate information.

Frontmatter Fields

For synthesized knowledge notes, include:

sources:
  - "[[daily/20260401-daily]]"
  - "[[meeting-notes/regulatory-review-20260328]]"
  - "Session transcript 2026-04-01T14:30"
synthesized: 2026-04-03
confidence: provisional
maturity: seedling
last_reviewed: 2026-04-03

Maturity levels (optional, tracks evidence strength):

• seedling — single source, provisional confidence. May not survive review.
• budding — corroborated by 2+ independent sources. Worth linking to.
• evergreen — reviewed, stable, established confidence. Core knowledge.

Inline Attribution

When a specific claim comes from a specific source, cite it inline:

• "Platform liability is shifting toward curation-based models ([[20260401-meeting-regulators]])"
• Use [[wikilinks]] for $ACA_DATA sources, markdown links for external sources

Rules

• Never synthesize without attribution — if you can't cite where a claim came from, don't assert it
• Distinguish observation from editorial — agents extract and synthesize but leave editorializing to the user
• Preserve uncertainty — use confidence levels. Don't upgrade provisional to established without additional evidence
• Source chain: When synthesizing from other synthesized notes, include the full chain (the intermediate synthesis AND its original sources)

Maps of Content (MOCs)

A Map of Content is a navigational hub note that curates links to related notes on a topic.

When to Create

Create a MOC when a topic area reaches a "mental squeeze point" — typically 5+ related notes that would benefit from a navigational index. MOCs are created by the /sleep consolidation cycle or manually.

Format

---
title: "MOC: Topic Name"
type: moc
tags: [moc, topic-area]
created: YYYY-MM-DD
last_reviewed: YYYY-MM-DD
---

Structure

MOCs contain curated links with brief annotations, grouped thematically:

# MOC: Platform Regulation

## Core Concepts

- [[platform-liability-frameworks]] — distinction between hosting and curation models
- [[content-moderation-at-scale]] — practical challenges of automated enforcement

## Australian Context

- [[osb-act-overview]] — Online Safety Bill structure and key provisions
- [[esafety-commissioner-powers]] — regulatory enforcement mechanisms

## Open Questions

- How will AI-generated content affect platform liability? (no settled answer yet)

Maintenance

• MOCs should be reviewed when the /sleep cycle detects they may be stale
• Add new notes to relevant MOCs when creating them
• Split MOCs that grow beyond ~30 entries

General Knowledge (Fast Path)

For factual observations NOT about the user. Location: knowledge/<topic>/

Constraints:

• Aim for concise notes (under 500 words for knowledge, under 200 for atomic facts)
• [[wikilinks]] on ALL proper nouns
• One fact per file for atomic knowledge; synthesized notes may cover a topic

Topics (use broadly):

• cyberlaw/ - copyright, defamation, privacy, AI ethics, platform law
• tech/ - protocols, standards, technical facts
• research/ - methodology, statistics, findings

Format:

---
title: Fact/Case Name
type: knowledge
topic: cyberlaw
source: Where learned
date: YYYY-MM-DD
---

[[Entity]] did X. Key point: Y. [[Person]] observes: "quote".

Background Capture

For non-blocking capture, spawn background agent:

Task(
  subagent_type="general-purpose", model="haiku",
  run_in_background=true,
  description="Remember: [summary]",
  prompt="Invoke Skill(skill='remember') to persist: [content]"
)

Output

Report both operations:

• File: [path]
• Memory: [hash]