trie

# <img src="landing/logo.svg" width="28" alt="trie logo" /> trie > **The self-hosted index of meaning and intent for your codebase.** [](LICENSE) [](https://www.python.org/downloads/) [](https://computerreinvention.com/trie/docs.html) **[Website](https://computerreinvention.com/trie/)** · **[Documentation](https://computerreinvention.com/trie/docs.html)** --- Every coding agent's session starts by re-deriving context: _what does this code do, and why is it like this?_ The code answers the first question bit-by-bit and the second not at all. Commit messages describe diffs, not decisions; documentation rots because nothing stops it from rotting. Symbol-level intent lives in the programmer's head or between an agent's conversation turns — and is lost either way. _trie_ makes this context explicit and permanent by creating and maintaining two separate indexes within your codebase. - **The Index of Meaning** (`triefacts`) — one Markdown file per source file, one section per symbol, describing what each thing does. LLM-generated, regenerated only when the source actually changes, cascade-aware, and verified at commit time. Entries in a triefact file have the same graph dependency as the actual source code. - **The Index of Intent** (`triediffs`) — the reason each symbol changed, recorded when the change is made and enforced at commit time. Archived per commit as a _triediff_ that shows up in your PRs. Months later, `trie read <symbol> --history` tells you the actual why. Also makes for a very nice and readable summary at PR time. Reduces review pressure drastically.  ## Quickstart Requires **Python 3.11+**, [`uv`](https://docs.astral.sh/uv/), and [`ripgrep`](https://github.com/BurntSushi/ripgrep) on PATH, plus an Anthropic API key in `ANTHROPIC_API_KEY` for prose generation (the gates, graph, and queries run offline). ```bash uv tool install git+https://github.com/computer-reinvention/trie cd /path/to/your/project export ANTHROPIC_API_KEY=... trie init # writes trie.toml, scans the symbol graph trie setup # wire up your coding agent (auto-detects opencode / claude-code / cursor / ...) trie sync # generate the meaning index ``` That's the loop: `init` → `setup` → edit → `trie sync` before each commit. A pre-commit hook will make sure the agent always records intent with `trie patch` for each symbol. The coding agent has to follow this structure by design - it can't commit otherwise. `trie setup` will override your agent's default `grep` with `trie grep` which is backed by the meaning and intent index - ~3.5 less round-trips. It will also add tools that invoke `trie read` and `trie trace` which utilize graph semantics over triefacts to one-shot the same questions that take a non-augmented agent many reads and greps. ## Documentation To learn more about how each index works, the commit gate, agent integration details, the command reference, configuration, costs, CI, team adoption, and troubleshooting check out the docs at **[computerreinvention.com/trie/docs.html](https://computerreinvention.com/trie/docs.html)**. ## License MIT. See [LICENSE](LICENSE). Contributions welcome — issues, discussions, PRs. For large changes, open an issue first so we can talk through the design.