Why your AI coding agent forgets team decisions (and what to store instead)

Many AI coding agents like Cursor and Claude Code forget team decisions when used together. The Claude.md tool only remembers decisions within a single session. Rules files remain static and don't carry team context. This causes repeated explanations of architecture tradeoffs and agents rediscovering past work. Using third-party memory services can be problematic without proper compliance. Built-in memory is usually limited to individual machines or repositories. Shared memory stores are often missing or divergent.

To store shared team context effectively, the article suggests persisting key information like decisions, reasons for those decisions, scope (which project or environment), and provenance (who made the decision and when). This can be stored in structured formats like Postgres databases. Decision-making conventions should be stored consistently, such as naming conventions, error handling guidelines, and known pitfalls.

Local memory works well for solo developers, but shared memory becomes valuable when multiple people work on the same codebase or switch between clients without a private store. Shared memory also helps with audit trails. While shared memory doesn't have to mean someone else's cloud, many teams find a local-first approach for individuals combined with an optional team store on their own infrastructure works best.

When designing memory access, keep the tool surface small with a few well-defined task-shaped operations like write, search, recall, and forget/supersede. Always include provenance information with results so users know when memory entries are stale or outdated. Avoid over-recalling too many memories in a single prompt, as this wastes context and confuses the model. Limit ownership of memory entries to prevent garbage accumulation, and consider light review processes for high-impact decisions.

Persisting full chat transcripts is ineffective for memory storage. Instead, focus on extracting key decisions and leaving the rest of the session intact. Be cautious about storing sensitive data like secrets or PII in agent memory; it's a policy issue rather than a feature. Finally, a simple habit of recording decisions succinctly (Decision: X. Reason: Y. Scope: Z.) can compound over time, regardless of the specific memory system used.

Palace, a local-first MCP memory layer, offers an open-core solution with optional self-hosted team functionality.