Open Brain
arXivPaperNeeds Review

CoReMem: Cost-Internalizing Memory for Successive Coding Tasks with Prospective Valuation

CoReMem tests repository memory by measuring whether stored task traces help on the next task. On ChainSWE, it improved a 27B actor’s resolve rate while reducing steps and token use.

arXiv · Nov 12, 2026
Open Source Open MarkdownOpen JSON
Source Summary

CoReMem selects compact repository memory by comparing task runs with and without trajectory segments, then adapts the agent based on measured memory value. Its memory centers on **repository locations** and **executable checks**.

Practical Implication

For builders running agents across successive tasks, this suggests keeping memory tied to actionable anchors and evaluating whether those anchors reduce work on later tasks.

Agent-Ready Context
CoReMem selects compact repository memory by comparing task runs with and without trajectory segments, then adapts the agent based on measured memory value. Its memory centers on **repository locations** and **executable checks**.

For builders running agents across successive tasks, this suggests keeping memory tied to actionable anchors and evaluating whether those anchors reduce work on later tasks.

On ChainSWE, the **27B actor** improved resolve rate from **21.0% to 25.7%**, with 17% fewer steps and 11% fewer cache-aware tokens. The result is from one benchmark; the abstract does not establish how well it transfers to other repositories or workflows.
Connected Context · Feed7 Judgment

CoReMem treats repository memory as a measured investment: it tests trajectory segments against later task outcomes and retains actionable locations and checks. Its ChainSWE result suggests this can help across successive coding tasks, but transfer beyond that benchmark remains open.

Context Map
agentcoding#agent-memory#harness-engineering
Uncertainty
On ChainSWE, the **27B actor** improved resolve rate from **21.0% to 25.7%**, with 17% fewer steps and 11% fewer cache-aware tokens. The result is from one benchmark; the abstract does not establish how well it transfers to other repositories or workflows.