How MemoraX Code Gives Coding Agents Long‑Term Memory to Stop Re‑Explaining Projects

Modern coding agents such as Claude Code and Codex can understand codebases, locate bugs, and even complete complex development tasks, yet when used over multiple days they repeatedly ask developers to re‑explain project history, because the agents lose accumulated project experience after each conversation.

MemoraX Code addresses this by adding a persistent memory layer that combines a local repository memory (to capture current code structure and historical evolution) with a cloud‑based long‑term memory that stores cross‑task, cross‑dialog, and cross‑agent project experience. Instead of feeding the entire history to the model each time, the system retrieves only the information that is truly relevant to the new task.

The authors point out that standard context‑window expansion and compaction merely preserve raw tokens, but do not decide which compressed information is worth keeping. MemoraX Code therefore augments the compression step with a learned judgment about what constitutes valuable experience, such as constraints like “module changes must remain compatible with the previous database version”.

To turn compressed memories into actionable knowledge, MemoraX Code automatically extracts Procedure Memory : it scans historical coding trajectories, identifies reusable analysis, implementation, and verification steps, and abstracts them into structured Skill objects. In a test set of 123 historical task fragments, the system distilled 15 concrete engineering insights and grouped them into four Procedure Memory categories.

Evaluation on the Agent Memory Leaderboard (AML) Coding Track shows MemoraX Code achieving a score of 62 , ranking first and improving the problem‑solving rate by 10 % over the previous state‑of‑the‑art Claude Mem solution.

In a separate three‑hour complex development task, using Procedure Memory raised the overall task score from 11.71 to 70.30**, increased the number of passed critical checks from 2/13 to 10/13**, and cut the estimated API‑call cost from $31.48 to $24.37 (a 22.6 %** reduction). The authors explain that the gain comes from avoiding repeated exploration of already‑solved sub‑problems.

Internal testing also shows that 85.5 %** of memory‑triggered actions align with user judgments, and 81.2 %** of recalled memories receive positive feedback, indicating that the system learns not only to remember but also to recall at the right time and with the right relevance.

MemoraX Code’s training pipeline treats memory as a reward‑driven capability: the model is rewarded when recalled information improves downstream decisions, and it continuously learns what to store, update, or discard. The architecture isolates the training infrastructure from private user data, ensuring privacy while enabling iterative improvement.

Support for four major coding agents (Codex, Claude Code, Deepseek Harness, OpenCode) is already available, with additional platforms in integration. The system also provides a management UI for users to view, edit, or delete stored memories.