Skip to Content
remimem

Limitations and workarounds

What remimem does not do well yet, and what to do about each one.

An honest list. Each item says what to do about it today.

Structural support covers four languages

TypeScript (and TSX), JavaScript, Python, and Go get symbols, a call graph, and per-function chunks. Every other language is searchable as text in 60-line windows, without symbols.

Workaround: text search still works, and vector search often finds the right window from a description. To add a language properly, see Adding a language.

No type inference in the call graph

user.save() is linked only when exactly one symbol in the repository is named save. In a hand-check, about one in four of these name-match links was wrong, and many real calls get no link.

Workaround: name-match links carry the lowest weight (0.2). To drop them, set "edgeWeights": { "name-match": 0 }. find_references may miss callers that go through untyped values; complement it with search_code for the name.

Documentation counts a quarter

Results from files that are not parsed as code (Markdown, JSON, SQL) are multiplied by 0.25, because changelogs and docs were crowding out code in the benchmark.

Workaround: if you mostly ask about prose, set "retrieval": { "textFileWeight": 1 }.

The benchmark tasks are easier than real bug reports

They are commit messages, written after the change, and they often name what changed. The test split is 36 tasks from three repositories, and the baseline is a simulation of a grep-driven agent.

Workaround: run the benchmark on your own expectations with --tune, and judge by remimem stats and how your agent behaves on your code.

Remote embedding providers are untested against live services

The OpenAI and Voyage providers are covered by tests against stubbed responses only.

Workaround: the default local model is the measured path. If you use a remote provider, run remimem doctor first: it makes one real embedding call and reports the result.

The install is large

The local embedding runtime brings about 430 MB into node_modules, and a first npx run downloads and unpacks about 750 MB into the npm cache. On a fast connection, install to first answer takes about 26 seconds; on a slow one, much longer.

Workaround: install once globally (npm install -g remimem) so npx does not re-resolve it, or keep a warm npm cache in CI.

The end-to-end agent benchmark has never been run

remimem bench --e2e is designed to give a real agent the same tasks with and without remimem. It costs money, so it only prints an estimate without --yes, and the agent flags it would use are unverified.

Workaround: none yet; the token benchmark is the measured evidence.

Memories cost a little code

With memories attached, the share of right files shown with a full body dropped by about two points in the benchmark (90% against 92%), because memory takes up to a quarter of the budget. Only a quarter of the memories attached there were relevant.

Workaround: keep memories few and specific, link them to code, and raise memoryMinSimilarity if irrelevant ones appear. "memoryLimit": 0 turns task-matched memories off; linked ones still show.

The default embedding model is not trained on code

A code-trained alternative found no more of the right files in the benchmark and indexed five times slower, so it is an option rather than the default. See Embedding providers.

Change detection trusts size and modification time

A change that keeps both is missed by the quick check before tool calls.

Workaround: remimem index reads and hashes every file.

Platform gaps

  • Vector search has no prebuilt extension for Windows on ARM.
  • TypeScript path aliases from tsconfig.json (paths, baseUrl) are not read: imports through them resolve only when exactly one file matches.
  • Go struct embedding and implicit interface satisfaction produce no graph links.
  • Outside a git repository, .gitignore is not read; only the built-in excludes apply. Use exclude in the config.
  • The index stores each vector twice (search and cache), about 1.5 KB extra per chunk. remimem index --full prunes unused cache entries.

On this page