Your First Insights

What Gets Extracted

ContextPool reads your chat transcripts and asks an LLM to extract the moments that are actually worth remembering. Not summaries of conversations — specific, actionable engineering insights.

Each insight has a type:

Type	What it captures
bug	A specific bug found — what it was, what caused it
fix	The solution applied — the exact change that resolved it
decision	An architectural or design choice made and why
pattern	A reusable approach or convention established
gotcha	A non-obvious pitfall that took time to discover

---

Example Output

Given a session where you debugged a Rust build failure, cxp might extract:

- **gotcha** reqwest needs `rustls-tls` feature on Alpine/ARM — default OpenSSL fails to link
- **decision** switched from `tokio::spawn` to `rayon` for CPU-bound extraction — avoids blocking the async runtime
- **fix** added `CARGO_NET_GIT_FETCH_WITH_CLI=true` to fix private crate auth in CI

These are stored in a .summary.md file and indexed by the MCP server. Your agent can retrieve them with get_project_context or find them by keyword with search_context.

---

Getting Higher Quality Insights

Have longer, more specific conversations. Short back-and-forth chats produce fewer insights. Sessions where you work through a real problem — debugging, designing, refactoring — tend to produce the most valuable memory.

Name things explicitly in chat. If you say "let's use a queue here instead of polling", that's a decision worth remembering. The more explicit the reasoning in your session, the better the extraction.

Run cxp init after sessions, not just before. Index new sessions regularly so your memory stays current. The MCP fetch_project_context tool does this automatically — but running cxp init manually from the CLI is faster for bulk historical imports.

---

What Gets Filtered Out

The extraction is aggressive about dropping noise:

• Tool call results and file contents injected by the IDE

• Generic explanations ("here's how async/await works...")

• Exploratory chat with no concrete outcome

• Thinking blocks and internal reasoning traces

• Lines longer than 500 characters (usually raw data)

Only what's worth telling a future engineer — or a future agent — makes it through.

---

Sharing With Your Team

Commit your ContextPool/ directory to git. Your teammates will automatically have your insights the next time they clone or pull:

git add ContextPool/
git commit -m "chore: add contextpool memory"

Or use Team Sync for automatic cloud sharing across your whole engineering team.