Agents are learning to remember — by dreaming, compressing, and governing what they keep

Dreaming over memory

Anthropic's new research on agent memory starts from a simple definition: memory is knowledge gained from experience that is retained between sessions and saved as readable context. It excludes changes to model weights, in-context learning, and compaction. The headline technique is what the researchers call dreaming, in which models use idle compute time to reorganize context, fix mistakes, and produce new notes. Anthropic said: "Dreaming gives Claude a way to test its notes against the original transcripts." m:…2227

In a preprint, the results are striking enough to take seriously. Building a wiki from a 350,000-file corpus improved performance on 25 questions from 13% to 48% over 128 sessions. Combining search and wiki access improved aggregate scores to 61%, compared with 53% for search alone. Prompting Claude to read memory files before answering nearly doubled the benchmark score for Sonnet 4.6, and Claude Opus 5 was the first model to close the learning loop on the CL-Bench database task. The counterintuitive finding is that for stronger models, giving them freedom to judge what is worth remembering worked better than setting fixed rules in advance. m:…2227

There is a limit, and it is instructive. In one test, Claude Sonnet 3.7 wrote 413 memory files while playing Pokemon Red, while Fable 5 wrote nine files and made better progress. More memory is not better memory. Anthropic's own Claude Code limits injected memory to the first 200 lines or 25 KB of its MEMORY.md file, which suggests the practical answer is curation, not accumulation. m:…2227

Stop reading tool output as text

A second preprint, from researchers behind a method called Coca-Cola, attacks the same problem from the other end: the tool output itself. LLM agents typically append raw tool outputs to their context, and long, noisy observations eat context and dilute attention. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space, so the agent reasons over a compressed representation instead of pages of text. The authors' own framing is that tool observations need to support the agent's decisions, and they do not have to remain readable text. m:…5718

In a preprint, the method cut agent token consumption by 70.01% to 89.24% and end-to-end runtime by 24.45% to 55.97% across four QA benchmarks and two model families, while consistently improving answer accuracy over uncompressed agents. The trick is Action-Preserving Training, which aligns the agent's next-action distributions under full and compressed observations, so the compressed version keeps the information that changes what the agent does next. A related preprint on offline retrieval-augmented generation found a parallel integrity gap: signed knowledge packs verify who published them but not whether the embedding vectors actually encode the associated text, and attacks on a Raspberry Pi 5 cut correct answers from 0.75 to 0.39. m:…5718 m:…5717

Govern what the agent remembers

The enterprise side of the same week is about governance, not algorithms. DataHub argues that agent context management works at two levels: the agent harness, which handles context per request, and the organizational context platform, which governs shared context per domain. Its example is Miro, which raised its analytics agent accuracy from under 40% to over 90% by adding expert-written context. DataHub reports that 83% of IT and data leaders agree agentic AI cannot reach production value without a context platform, and that context rot sets in within a single run as the window fills, while stale context builds up across agents over time. m:…2184

Squirro makes the same point from the knowledge side: context layers fail because they rely on documentation and ignore tacit knowledge in chat threads, tickets, and expert experience. Its capture process distills threads, finds gaps, asks the expert, and governs the result, with editor approval before an answer enters the context layer. As Squirro's Jan Overney put it: "An AI agent has nobody to walk over to." Gartner named AI context platforms a top strategic technology trend for 2027, and projects that prioritizing semantics in AI-ready data could raise agentic AI accuracy by up to 80% and cut costs by up to 60% by 2027. Only 22% of enterprises currently capture significant value from generative AI, which suggests the memory problem is also the adoption problem. m:…2508