fastest way to get knowledge and memory for your Claude Code sessions
flex compiles your Claude Code session history for vector and hybrid retrieval. your AI agent connects via MCP, discovers the schema at runtime, and writes SQL against your history.
install:
curl -sSL https://getflex.dev/install.sh | bash -s -- claude-codemost memory systems start working after you install them.
flex works retroactively — as soon as you install, all of your sessions become queryable via the MCP tool. you can ask claude how you installed the cloudflare tunnel yesterday, why you made a specific edit to your landing page, and much more.
retrieval tools typically have minimal metadata to filter on.
flex captures everything: the exact session, whether it spawned agents, what files and repos it touched, which project it belongs to. filtering happens before scoring — the vector engine only scores what survives. on top of that, flex builds knowledge graphs automatically — community structure, hub sessions, file co-edit patterns, project attribution.
vector search typically surfaces similar content and stops there.
flex lets you suppress a topic, weight by recency, diversify across subtopics, trace a direction through embedding space — all composable in one query. if you want "architecture documents" but not "changelogs", just ask claude; he'll use the similar:architecture and suppress:changelog tokens.
flex tracks every session's messages and operations. get information on what session created a file, what prompts were used to create it, and why it was changed last. just ask:
"Use flex: what's the history of worker.py?"
the history of worker.py is followed beyond file moves and renames. never lose track of your reasoning.
the hardest question in software: why was it done this way? flex finds the session where the decision happened and reconstructs the path — which approaches were considered, which failed, and why you landed here. just ask:
"Use flex: how did we create the curl install script?"
this works for technical questions, practical concerns and logistical setups as well. makes it easy to grab the runbook for setting up your website from three weeks ago.
sessions are grouped by the projects they touch, the files that got the most edits, and key decisions. just ask:
"Use flex: what did we build this week?"
claude runs a series of queries behind the scenes and comes back with an overview of your recent week.
semantic search just surfaces similar content. flex can search for one topic while suppressing another. just ask:
"Use flex: 5 things we talked about this week outside the main project"
it can also get a diverse sample of information.
your knowledge base is one file on your machine.
ls ~/.flex/cells/
claude_code.db 284M
# back it up
rsync -av ~/.flex/cells/ backup:~/
# query it directly — open format
sqlite3 claude_code.db "SELECT COUNT(*) FROM sessions"
4547everything runs in-process. no external services, no cloud dependency.
query from the terminal without claude:
flex search "@digest days=3"
flex search "@file path=src/worker.py"
flex search "SELECT COUNT(*) FROM sessions WHERE project = 'myapp'"same queries, no MCP required.
curl -sSL https://getflex.dev/install.sh | bash -s -- claude-codeMIT · Python 3.12 · SQLite · getflex.dev · x