I gave my AI coding agents a local long-term memory layer — 8 things that broke

Every new coding session starts from scratch, requiring agents like Claude Code and Cursor to remember why the task was approached a certain way. The agent struggled to recall past decisions, flaky tests, and configuration flags. To address this, a local long-term memory layer was implemented with four key responsibilities: recording, retrieving, consolidating, and pulling project files as raw material. However, several issues arose during implementation:

1. The memory service running independently from the agent caused failures when the service was not running. Treating the transport process as stateless and disposable helped mitigate this issue.

2. Storing everything indiscriminately led to decreased output quality, as irrelevant memories wasted tokens and mislead the agent. A relevance signal became necessary for recall quality.

3. Running multiple agents concurrently resulted in interleaved and contradictory records. Implementing a single-writer discipline with locking resolved this problem.

4. Memories went stale and contradicted each other after a few weeks, necessitating an explicit consolidation pass to detect conflicts and supersede outdated information.

5. Recall latency added visible delay, which could be reduced by maintaining an index and imposing a hard time budget on recall.

6. Secret management became a concern when project files were ingested, as it could expose sensitive information. Filtering sensitive data at the ingestion boundary prevented such issues.

7. Deciding which files carried durable context proved challenging, requiring proper tuning of the file-bridge to distinguish between durable context and generated noise.

8. Evaluating the effectiveness of the memory layer was difficult, as no clear metrics were established. Implementing a replay set of past tasks helped measure the impact of the memory layer.

The resulting product, HyperMarrow, is a local long-term memory layer for coding agents that addresses these eight failure modes. It consists of four modules and provides four ways to connect: MCP server, direct desktop integration, Python SDK, and Node SDK. HyperMarrow runs on Windows and offers a 30-day full-feature trial with WeChat QR or SMS sign-in. The checklist of transferable failure modes is provided for those who wish to build their own solution.