What is repo memory for AI coding agents?

Repo memory is persistent, repo-specific evidence an agent can retrieve before editing: central files, symbol definitions, callers, tests, dependencies, diffs, and prior decisions tied to git state.

How is repo memory different from chat memory?

Chat memory remembers conversation context. Repo memory is grounded in the current codebase and git history, so the agent can verify names, callers, and decisions against files that actually exist.

When is Sverklo the wrong tool?

If you know the exact string, grep or ripgrep is usually faster. Sverklo is for relationship-heavy edits where callers, tests, dependencies, and prior decisions matter.

Repo memory MCP

Give your coding agent repo memory before it edits.

Sverklo is a local-first MCP server that gives Claude Code, Cursor, Windsurf, Codex CLI, and other agents grounded evidence from your repo: files, symbols, callers, tests, dependencies, diffs, and decisions pinned to git state.

Proof first: run the no-write receipt in any repo before installing MCP config.

npm exec --yes --package=sverklo@latest -- sverklo prove --no-write --guided --markdown

The receipt should show a meaningful center of the repo and at least one real relationship. If it does not, that is a useful correction.

What repo memory means

Chat memory remembers what happened in a conversation. Repo memory is different: it is retrieved from the codebase and tied to the current git state, so an agent can verify what actually exists before it edits.

Symbolslookup resolves names to definitions with file and line evidence.

Callersrefs and impact show who uses the thing about to change.

Diffsreview_diff maps changed symbols to tests, risk, and similar code.

Decisionsremember and recall keep project decisions attached to git SHAs.

Verificationverify checks that cited evidence still points at the same code.

SearchHybrid BM25, local ONNX embeddings, PageRank, and structural parsing run on your machine.

Where it helps

Refactors where the agent needs callers and tests before touching code.
Large repos where the obvious files are not the only files that matter.
Long-running work where prior decisions get lost after context compaction.
Code review flows where changed-symbol risk matters more than a generic diff summary.

Where grep still wins

If you know the exact string, use grep or ripgrep. Sverklo is for the next step: relationship-heavy questions where the literal name is not enough, or where the agent needs proof before it writes code.

How to try it

Start with the no-write proof guide. If the receipt is useful, continue with install or the Codex CLI recipe. If it is wrong, share the failure in the proof thread.