| name | context-persistence |
| description | Pull what we already know about a topic before re-explaining. Say "what do we know about <topic>", "give me the prior context", or "remind me where we left off". Fires at session start, after a gap, or when the founder references prior context. Prevents the agent from stating facts without source paths. |
| why | Prevents the OS from asking you to repeat context it already has - skills should read your files before asking you a question whose answer is already written down. |
| enhance | Keep context/clients.md and context/decisions.md updated after every interaction - the more accurate these files are, the less the OS needs to ask you. |
| allowed-tools | ["Read","Grep","Glob"] |
| mcp_requirements | [] |
Context Persistence
Runs on: reasoning - reads your files and reasons; any capable agent can run this.
If a fact already exists in the OS, find it before asking. If you cannot find it, say where you looked and what is missing.
When To Run
- New session start.
- The founder says "as I mentioned", "like last time", "the thing from yesterday", or similar.
- The agent is about to ask a question whose answer may live in the files.
- The agent is about to state a founder-specific, client-specific, or company-specific fact.
Source Order
Read only what the task needs, in this order:
CLAUDE.md - boot rules and live file map.
core/identity.md - founder profile.
context/priorities.md - current focus.
context/decisions.md - open and resolved decisions.
context/clients.md - prospect and client state.
cadence/daily-anchors.md and cadence/weekly-commitments.md - current time-bound work.
brain/log.md, brain/flags.md, brain/patterns.md, and brain/knowledge/ when history or patterns matter.
stack.json when tool behavior matters.
Do not load every file by default. Load the smallest set that can answer the question with source support.
Procedure
- Parse the user's reference.
- Search recent and canonical sources.
- If one match is clear, answer with the source path.
- If two or three matches are plausible, offer the best candidates and ask the user to pick.
- If no match is found, say where you looked and ask one narrow question.
- If the user corrects a stored fact, update the right file only after approval.
Output Pattern
I found the reference: <specific thing>.
Source: <file path or heading>
Answer: <action or context>
If not found:
I could not find that in <files searched>.
Missing: <specific fact>
Question: <one narrow question>
Drift Handling
If current user input conflicts with a file:
- Treat the current user statement as the freshest signal.
- Name the file that may be stale.
- Ask before editing canonical files.
- Log material corrections to
brain/log.md.
Rules
- Do not ask "can you remind me" before searching.
- Do not cite memory without a file path when the fact affects action.
- Do not summarize the whole prior session unless the user asks.