Agent Memory Lifted Opus 5 From 0.564 to 0.828, If the Agent Reads It

Memory design decides whether an agent improves with use, and how the model uses its memory matters as much as the store. Anthropic’s Lance Martin defines agent memory as knowledge gained from experience and kept across sessions, excluding weight updates, in-context learning and compaction. On the CL-Bench 1.0 database task, adding memory lifted Opus 5 from 0.564 to 0.828 mean reward and Opus 4.7 from 0.383 to 0.700. Sonnet 4.6 moved only from 0.330 to 0.364, until it was prompted to read and test its memory, which raised it to 0.641.

Volume is a poor proxy for quality. In a Pokémon Red evaluation, Sonnet 3.7 wrote 413 memory files and earned one badge, while Fable 5 wrote nine files and earned four. Instructions change what an agent keeps: telling Opus 5 to record only failures cut its gain by 0.234. On a corpus of roughly 420 million tokens, a wiki built over 128 sessions raised wiki-only scores from 13% to 48%, and wiki plus search reached 61% versus 53% for search alone. The post’s proposed mechanism is “dreaming,” an offline pass in which Claude reviews its logs, fixes mistakes, writes notes and builds memory structures.

Salesforce reports the same pressure in production. Its team says it doubled Agentforce Memory’s performance by changing the memory policy rather than the architecture: an update now counts as a correction to existing knowledge. In Salesforce’s example, a customer who moves from Bangalore to Pune is no longer answered from the old address. Against the policy it replaced, recall rose 1.8 times for facts customers state outright and 2.2 times for facts that change over time. Those results are self-reported.

The takeaway for leaders funding agent projects: ship memory as a designed layer with its own evaluation. Decide what gets stored, keep the injected layer small, and test whether the model actually reads it before counting the outcome.