| name | memory-wiki |
| description | Use when a conversation, task, or channel message contains durable facts, preferences, decisions, or project context that should be preserved in the agent's markdown-first memory wiki. |
Purpose
Use this skill when you need to preserve useful memory without asking a human to learn bridge memory commands.
The human should speak naturally. You decide whether the information deserves memory, then update the wiki yourself.
Source Of Truth
- Markdown wiki files are source of truth.
- Raw capture files are inputs.
- SQLite indexes are derived helpers.
Default Path
Prefer the one-step command unless you specifically need separate stages:
~/.agent-bridge/agent-bridge memory remember --agent "$BRIDGE_AGENT_ID" --user <user-id> --source chat --text "..." --kind user
This will:
- create a raw capture
- ingest it into the active user's daily memory
- optionally promote it into long-term memory
Raw-Source Ingest Workflow
Use the staged flow when the source is long, noisy, or arrives as a file rather than a short chat message.
Examples:
- a pasted meeting transcript
- a Discord export or support thread
- a markdown note captured from another system
- a long email or incident write-up
Recommended flow:
- capture the raw source into
raw/captures/inbox/
- ingest it into the correct user's daily memory
- only then promote durable facts into curated pages
Example with a file source:
~/.agent-bridge/agent-bridge memory capture \
--agent "$BRIDGE_AGENT_ID" \
--user <user-id> \
--source notes-import \
--title "Weekly ops retro" \
--text-file /path/to/source.md
~/.agent-bridge/agent-bridge memory ingest \
--agent "$BRIDGE_AGENT_ID" \
--latest
If the source contains only a few durable facts, you may still prefer memory remember.
If the source is ambiguous or partly unverified, prefer capture -> ingest first and promote later.
When To Remember
Use memory when a message contains:
- stable preferences
- recurring workflow expectations
- durable personal context
- long-lived project context
- decisions that should affect future behavior
Do not remember:
- transient chatter with no future value
- facts that clearly belong only to the current turn
- speculative or unverified claims you would not want to rely on later
Which Kind To Use
--kind user
- default for one human's preferences, habits, constraints, or stable context
--kind shared
- facts that apply across humans for this agent
--kind project
- project/domain context worth keeping on a dedicated page
--kind decision
- policy changes, agreed rules, architectural choices
--kind none
- capture and ingest only, with no long-term promotion yet
Multi-User Guardrail
If the agent supports multiple humans:
- always identify the active user first
- write user-specific facts under that user's partition
- never mix one person's preferences into another person's memory
When unsure, prefer:
- daily user memory first
- then promote later after more confidence
Retrieval
Before asking the human to repeat known context, check memory:
~/.agent-bridge/agent-bridge memory query --agent "$BRIDGE_AGENT_ID" --user <user-id> --query "..."
If the derived index is missing or you want broader markdown-first matching:
~/.agent-bridge/agent-bridge memory search --agent "$BRIDGE_AGENT_ID" --user <user-id> --query "..."
Maintenance
Use these when needed:
memory lint
- find missing structure, pending captures, and wiki hygiene problems
memory rebuild-index
- rebuild the derived SQLite index after significant edits
Example
The human says:
I prefer weekly summary digests on Fridays.
You may preserve it with:
~/.agent-bridge/agent-bridge memory remember \
--agent "$BRIDGE_AGENT_ID" \
--user owner \
--source chat \
--text "The user prefers weekly summary digests on Fridays." \
--kind user