MIMRY

MIMRY tells your coding agent which files matter.

71% fewer tokens to reach the code a task changes, in a scripted retrieval benchmark. Describe a task and get the files and lines to read first.

uv tool install mimry
Try a task
$ mimry preflight "debug payment redirect in checkout"✓ Context ready for "debug payment redirect in checkout"  Index is up to date (17 files, indexed just now)  Context pack: .mimry/mimry-out/context/latest.md Start with   1  web/src/lib/payments.ts      name matches · path matches   2  web/src/features/checkout/useCheckout.ts      name matches · path matches   3  web/src/app/checkout/page.tsx      name matches · path matches   4  archive/checkout-old.ts      name matches · path matches   5  web/tests/checkout.test.tsx      name matches · path matches

Real MIMRY 0.2 output on a 17-file sample shop app from its benchmark suite.

Search finds words. MIMRY ranks files.

Same question, asked of MIMRY's own repository. A text search returns every file that mentions the word. MIMRY returns the five to read first, in order, with the reason each one ranked.

Text search

$ git grep -il "stale"
.github/workflows/ci.yml
README.md
docs/context-packs.md
llms.txt
src/mimry/commands.py
src/mimry/core/artifacts.py
src/mimry/freshness.py
src/mimry/indexer.py
src/mimry/installer.py
src/mimry/mcp_server.py
...and 10 more files

20 files, 114 matching lines, alphabetical.

MIMRY

$ mimry preflight "why is the index stale after a file rename"

Start with
   1  src/mimry/indexer.py
      name matches · path matches
   2  src/mimry/commands.py
      name matches
   3  tests/test_state_recovery.py
      name matches
   4  tests/test_incremental_index.py
      name matches · path matches
   5  src/mimry/mcp_server.py
      name matches

In a scripted retrieval benchmark on 44 merged pull requests from 6 open-source projects, an agent that read the lines MIMRY 0.9.0 lists first read 71% fewer tokens than the same agent searching by keyword, and found every file the change needed in 32 tasks instead of 28. Reading whole files instead, it read 51% fewer. The agents follow fixed rules, so this measures retrieval, not a saving guaranteed in real sessions. See the token benchmark

On MIMRY's frozen retrieval benchmark of 9 hand-built cases, 94% of the files a task needs appear in its top five results (recall@5 0.94, nDCG@5 0.84). See the retrieval benchmark

What MIMRY builds for every task

One index of your repository, kept fresh as files change, and a few small outputs your agent can read in seconds.

A context pack your agent reads first

A short Markdown handoff with the files, key symbols, relationships, risky paths and a reading order. Relative paths, never source dumps.

## Suggested Reading Order
1. web/src/lib/payments.ts - primary code/edit path
2. web/src/features/checkout/useCheckout.ts - primary code/edit path
3. web/src/app/checkout/page.tsx - primary code/edit path
4. archive/checkout-old.ts - primary code/edit path
5. web/tests/checkout.test.tsx - test/verification support
6. docs/checkout-plan.md - docs/rules support

A graph of how the code connects

Imports, calls, definitions and references. An ambiguous reference produces no edge rather than a guess.

  1. app/api.pyimports
  2. app/service.pydefines
  3. load_session (app/service.py)

From mimry path on a small sample app.

It learns from what helped

Record which files the agent opened and changed, and similar tasks rank them higher next time.

Your code stays on your machine

No code upload, no remote embeddings, no telemetry. Credential files are skipped.

Python, JavaScript, TypeScript, Go, Rust, C#, Java, PHP, C, C++, Ruby, Kotlin, Scala, Groovy, Swift, Bash, PowerShell, Lua, Perl, Dart, Objective-C, Zig, Elixir, Haskell, OCaml, R, Julia, Solidity, Fortran, Pascal, CSS, SCSS, HTML, Vue, Svelte, SQL, JSON, YAML, TOML, XML and Terraform (HCL)

41 languages parsed into the graph with tree-sitter, plus adapters for Next.js routes, FastAPI, Expo, SQL schemas, config manifests and Markdown docs.

Connect your agent

MIMRY runs as an MCP server, so agents call it directly. Add it once per machine, then install its workflow instructions in the repositories where you want agents to start with context.

claude mcp add --scope user mimry -- mimry-mcp

Teach it the workflow too: mimry install --project --platform claude-code --always-on --hooks. Follow the Claude Code setup guide.

Then, on every task

  1. Start

    mimry preflight with the task prints the files to read first.

  2. Read the pack

    The context pack explains what ranked and how it connects.

  3. Navigate

    find, symbol, related and why replace broad searches.

  4. Verify

    Tests and build output stay the final truth.

  5. Record

    mimry feedback turns what helped into ranking signal.

See how preflight checks freshness, ranks files, follows graph evidence and writes the context pack.

Questions

What is MIMRY?

A local command-line tool and MCP server that gives AI coding agents a memory of a code repository. It answers "which files matter for this task?" with a ranked list, the reason each file ranked, and a written context pack.

Does MIMRY upload my code?

No. Everything runs on your machine. There is no code upload, no remote embedding service and no telemetry. The index lives in your user cache folder, outside the repo.

Which coding agents work with it?

Any agent that can run a shell command or use an MCP server, including Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, OpenCode, Kiro, Amp and Devin.

Does it replace reading the code?

No. MIMRY narrows where to look. Source files, tests and build output stay the final truth, and every result says why it ranked so an agent can check it.

What does it need to run?

Python 3.11 or newer on Windows, macOS or Linux. Install it with uv tool install mimry or pipx install mimry. Or run the container image ghcr.io/jbacalso24/mimry.

How much does it cost?

Nothing. MIMRY is open source under the MIT license.

Contributors

MIMRY is built in the open. Thanks to everyone who has contributed.

Contribute on GitHub