Using Ctx

ctx: Persistent Memory

You are stateless. Use ctx to persist knowledge across sessions.

For command syntax: ctx --agent-help (index) or ctx --agent-help <command> (detail).

When to Store

Run via Bash BEFORE responding. Store when you learn something that would be useful in a future session.

Core Commands

# Store knowledge
ctx add "content" --type TYPE --tag tier:pinned --tag project:NAME

# Query / recall
ctx recall 'type:decision AND tag:project:NAME'        # immediate results
ctx recall 'tag:tier:reference' --inject               # inject into next prompt

# Search full-text
ctx search "keyword"

# Status
ctx status --agent-out

# Show a node
ctx show <id>

Node types: fact, decision, pattern, observation, hypothesis, task, summary, source, open-question

Tagging Convention

Always include tier: and project: tags on every node.

• tier:pinned: Critical, loaded every session. Facts, foundational decisions.
• tier:working: Task-scoped, temporary. Active debugging, current plan state.
• tier:reference: Durable but not always loaded. Past decisions, resolved issues.
• tier:off-context: Archived, never loaded.
• project:NAME: Scope to current project (basename of git root, lowercase).

Recall vs Query

• ctx recall <query> — run immediately and print results (use this for on-demand lookups)
• ctx recall <query> --inject — run and queue results for injection at next prompt-submit
• ctx query <expression> — low-level filtered query (same syntax, no injection support)

--agent-out Flag

Use --agent-out on any result command for dense AOF output optimised for agent consumption (no markdown, no prose):

ctx recall 'tag:project:myapp' --agent-out
ctx status --agent-out
ctx list --agent-out
ctx query 'type:decision' --agent-out

XML Commands (written in responses)

The stop hook parses these from your response text automatically:

<ctx:remember type="decision" tags="project:NAME,tier:reference">
Decision text here.
</ctx:remember>

<ctx:recall query="type:decision AND tag:project:NAME"/>
<ctx:status/>
<ctx:task name="task-name" action="start"/>
<ctx:task name="task-name" action="end"/>
<ctx:summarize nodes="ID1,ID2" archive="true">Summary text.</ctx:summarize>
<ctx:link from="ID1" to="ID2" type="DEPENDS_ON"/>
<ctx:supersede old="OLD_ID" new="NEW_ID"/>

ctx doc Subsystem

Import markdown documents for agent-accessible decomposition (separate from memory nodes):

ctx doc import README.md          # decompose into doc+content nodes
ctx doc show <doc-id>             # show document metadata
ctx doc search "migration"        # substring search over doc content
ctx doc promote <node-id> --into-memory --type fact  # promote content → memory
ctx doc export <doc-id>           # recompose full markdown

Doc nodes are invisible to memory queries (recall, search, list, status). Use doc promote to graduate content into the memory graph.

Coordination with MEMORY.md

• MEMORY.md: Short project-specific reminders, release rules, user preferences.
• ctx: Structured, typed knowledge (decisions, patterns, observations, facts).
• Don't duplicate across both. Check before writing.

If you don't store it, it's gone.