From repeated rediscovery to reusable experience.
A coding agent discovers a repository convention, a test command, or the reason behind an architectural choice. A later session starts without that context and spends another turn finding it—or chooses a path the team already rejected.
UniMem can retain atomic facts and reviewed lessons outside the current context window. A fact records a proposition. A lesson expresses reusable guidance with an applicability condition. Neither should become permanent instruction merely because an agent once wrote it.
Start with a narrowly scoped project store. Keep instructions tied to the repository and task family they concern, and review changes before allowing them to become recurring guidance. This scenario illustrates an integration pattern, not measured customer results.
Attach to the tools you use.
The repository includes configuration targets for Claude Code, Cursor, Windsurf, Codex, Copilot, Gemini, Devin, OpenCode, and generic MCP clients. The local MCP server exposes memory operations without requiring the agent to run a shell command for every interaction.
Install into a disposable project first and inspect the generated configuration. Client-specific permissions, hooks, and invocation behavior still belong to the client’s environment. A supported installer target does not imply a commercial partnership.
Use an existing project fact as your first recall check, then make the agent’s use of the delivered context observable before introducing a memory experiment.
unimem init --dest ./mem
unimem install --target cursor
# Other supported target: codex, claude-code, windsurf
A convention can help—and later become stale.
Suppose a lesson tells the agent how to run a test suite. After a tool migration, that once-useful instruction may cause repeat failures. Persistent storage alone will not identify the change.
On eligible, low-risk tasks, a controlled memory trial can compare results with and without the lesson. Report an independently verified result, such as the agreed test-and-review outcome. A single successful run does not establish an effect.
Use evidence to inform withdrawal or a reviewed update. Keep critical security rules outside ordinary withholding. The newer causal gate can use context labels, but deployment support and the experimental assumptions need review before relying on it.
Memory measurementPut experience to work.
Try the local loop, or bring one workflow to a pilot conversation.
Technical details link to the inspected source. Check deployment requirements before enabling optional features.