# AI Agents Can Dream Now. You Still Need to Tuck Them In Source: https://robertpeake.substack.com/p/ai-agents-can-dream-now-you-still Kiwi News: https://news.kiwistand.com/stories/AI-Agents-Can-Dream-Now-You-Still-Need-to-Tuck-Them-In?index=0x6afa0d43fa5eb3319dd93d210791afed2da0ffbaeefdb388d71daf30f3c725795c2d6614 Submitted by misha on 2026-11-15 UTC, 1 upvote, 0 comments ## Summary The author argues that AI agents need deliberate curation of their context, such as end-of-session debriefs and carefully managed notes, rather than relying on automatic "dreaming" or always-loaded instruction files. Anthropic's dreaming process, in which Claude rereads session logs offline and builds them into a wiki, raises privacy risks, since transcripts can hold sensitive client details. Its evaluations also check results against session transcripts, which cannot capture the unspoken common ground of most knowledge work. Citing a study of four agent-model combinations, the author notes that repository instruction files did not significantly improve coding success and raised inference costs by 20–23%, though the author says this applies mainly to code. ## Article A colleague who has memorised the internet walks in, eager and sharp. They have never met you, never seen the project, and have no idea what has already been done. No surprise that bringing them in by the truckload seems to have created nearly as many problems as it has solved. No, I’m not talking about consultants. I’m talking about AI agents. The trouble is not how clever the agent is, but how hard it is to “get on the same page”, and “stay on the same page” with them over time. Without useful context carried over from last time, each session is a first day, and a new HR orientation. Many organisations have tried to solve this by pointing agents at data they have to hand: the organisation’s rules and structures, a sea of documents and databases. Even with judicious retrieval of this information, it can still quickly fill the agents’ “head” with too much to process. It accumulates day after day, and research shows that longer context can degrade performance even when retrieval is perfect (Du and colleagues, 2025). Most carbon-based life forms have a remarkable answer to the problem of drinking from the daily firehose of information pointed at us in the 21st century. Lately, so do some silicon-based ones. This is where the mischief starts. Complementary learning systems theory holds that the brain learns with two systems running at different speeds. The hippocampus rapidly encodes the day’s experiences, while the neocortex integrates them more gradually, reactivating those memories during sleep and quiet rest, and weaving new learning together with what we already know (McClelland, McNaughton and O’Reilly, 1995). Two details of the process matter for helping our little amnesiac AI friends. First, forgetting can be an active process, not merely an accidental loss, and it can help us by clearing out detail that no longer matters (Richards and Frankland, 2017). Second, consolidation turns knowledge into building blocks. Rats that had already learned a stable set of flavour–place pairings could add new ones more quickly (Tse and colleagues, 2007). What you already know well makes new learning cheaper, because you learn not just facts but patterns. Sleep, in other words, is a moonlight groundskeeper, pruning our mind’s garden in service to its health and beauty. Anthropic recently published an account of what it calls dreaming, building on earlier work on “sleep-time compute” (Lin and colleagues, 2025): an offline pass, inspired by the hippocampus, in which Claude rereads its own session logs, corrects its notes, and builds them into something closer to a wiki (Martin, 2026). The results are impressive in the lab. Yet a machine that rewrites its memories while you sleep deserves more careful consideration, lest it drift into nightmare. The first nightmare is about privacy. To dream, the model rereads the transcripts, and the transcripts can hold everything: the client’s numbers, the aside about a colleague, the gripe about your boss. Whatever the dream judges salient becomes a note that outlives the conversation. The second nightmare is about grounding. Look at what the dreaming is checked against. Every task in Anthropic’s article comes with a reference for correctness: a graded reward, a badge, a source document, and above all the session transcript, which the authors describe as a ground truth to validate memory against. That is necessary for such study, yet it also narrows the study to what can be checked objectively. In most knowledge work, the things that matter never reach the transcript: why we chose the more costly supplier, which client hates garlic, the strategy we shredded around the water cooler. Linguists call the shared understanding beneath a conversation “common ground”, and the act of creating it “grounding”. It is built jointly, one party offering and the other extending themselves into it (Clark and Brennan, 1991). Language models are notably poor at the extending part: in the conversations studied, they asked for clarification about a third as often as people did (Shaikh and colleagues, 2025). A model can reread a transcript on its own. It cannot ground itself in what was never said. The third nightmare is stuck in the rut. A misunderstanding, written down, becomes a note. Reread every night, the note becomes a rule, and a rule loaded into every session becomes the way things are done. Where nothing contradicts it, the gist can quietly harden into fact. Human memory keeps the gist and fills in details to fit (Payne and colleagues, 2009). And when language models are nudged toward a wrong answer, they often produce plausible explanations justifying it, without mentioning what actually swayed them (Turpin and colleagues, 2023). None of this means agents should not remember, or dream. It means somebody still needs to keep an eye on them while they are awake. Many coding agents read a standing instruction file at the start of every session: a CLAUDE.md, an AGENTS.md, a folder of “rules”. Write your lessons there and every session begins with them automatically, which looks very much like the amnesia problem solved. Yet the research on “context engineering” is sobering. One team tested four agent–model combinations on software tasks, with and without repository instruction files. Neither machine-written nor developer-written files significantly improved success over having no file at all, and machine-written files raised average inference costs by 20–23% (Gloaguen and colleagues, 2026). But before you go deleting your CLAUDE.md, as some articles suggest, remember this applies to code, which has been engineered over time to be among the most verifiable of complex outputs. That fact that agents can often get what they need straight from the code and code comments, without the English-language commentary you would write to orient a junior programmer, is hardly proof that curating context in knowledge work more broadly is dead. If anything, it illustrates the “jagged frontier” researchers have described (Dell’Acqua and colleagues, 2026), where coding at the far edge has been radically transformed, and the average office worker with a Copilot licence gets smart-sounding advice alongside face-palm-dumb stuff, and wonders, “What’s the big deal?” Stigmergy is a term Pierre-Paul Grassé coined in 1959 for termites building nests: each worker responds to the state of a half-built structure left by a colleague. AI sessions already work like this, each one reading the files the last one left behind. Here’s what sometimes happens: army ants can become trapped in circular trails, each following the one in front, sometimes circling until many lie dead or dying (Schneirla, 1944). Nobody was there to point out the ring they were wearing into a rut. I use four principles for curating context in knowledge work, and I credit them in part for results my peers find surprising. They are simple rituals that create grounding between me and my agents, and stop us going in circles. I perform all of them with the agent, so the LLM writes the notes in its own words. Close out at the end of every session Capture what was learned, and what was decided, from the session itself, while it is fresh. This helps you reallocate your own attention. It also stops your agents re-suggesting something you have already decided against over and over again. Move knowledge into the right place In my world there are four places new information can go for next time: the working standards every session reads, a reference library by topic, the folder of the project it belongs to, and templates for repeatable work. This is also the stage where I can redact sensitive information. This parcels information into documents and folders for auditing and inspection (by me or scheduled agents), and helps the agent load context just-in-time from a wiki-like index of options which, unlike dreaming structures, are at least partly aligned with how I think as well. Keep working standards succinct The always-read list can grow, but every addition needs to be succinct and meaningful, because every future session loads it in, which is not without cost. Store what’s not otherwise discoverable This is the golden rule in a world of rapid gains from reinforcement learning on pre-trained models. They can often pick up what they need from the code itself. Yet an agent has no idea who needs to be kept informed of the project status, unless you tell it. Without essential context, it can be both brilliant and dumb. The principles above keep what should be private out of what is shared, add what no transcript or codebase holds, and catch a misunderstanding before it wears into a rut. Equally important, they give me a ritual with my agents that ensures I keep track of their “understanding”, and so can manage them consistently toward good results over time. Notice what all this does not need: bigger models, more data, or esoteric smarts. It needs someone to run a decent end-of-project debrief, decide what is worth keeping, redact private information, and retire the ideas that no longer make sense. Those are management skills. They are the skills of onboarding, handover and retrospective review, and they can be taught. In an organisation using AI well, individual contributors become managers of agents, though not quite in the same way they would manage direct reports. No matter how good the models get, organisations developing “agentic management” skills will compound their investment in learning, while everyone else reintroduces themselves to AI every morning, or worse, to a rogue version fresh from its nightmare, ready to cause trouble. Your AI agents can dream now. Whether that is a good night or a nightmare depends on who decides what they get during the day. Put them to bed with care, and you can get a good night’s sleep, too, and pick up where you left off in the morning. If you’re interested in an automated way to invest in good context, my workspace kit is an experimental approach to supporting human-AI grounding rituals and knowledge promotion taxonomies. It is free and open source on GitHub: https://github.com/cyberscribe/agentic-workspace-kit Show moreShow less - Clark, H. H. & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine & S. D. Teasley (Eds.), *Perspectives on Socially Shared Cognition*, 127–149. American Psychological Association. - Dell’Acqua, F. et al. (2026). Navigating the jagged technological frontier. *Organization Science*. https://doi.org/10.1287/orsc.2025.21838 - Du, Y. et al. (2025). Context length alone hurts LLM performance despite perfect retrieval. *Findings of EMNLP 2025*, 23281–23298. https://doi.org/10.18653/v1/2025.findings-emnlp.1264 - Gloaguen, T., Mündler-Sasahara, N., Müller, M. N., Raychev, V. & Vechev, M. (2026). Evaluating AGENTS.md: Are repository-level context files helpful for coding agents? arXiv:2602.11988v3. https://arxiv.org/abs/2602.11988v3 - Grassé, P.-P. (1959). La reconstruction du nid et les coordinations interindividuelles chez *Bellicositermes natalensis* et *Cubitermes* sp. La théorie de la stigmergie. *Insectes Sociaux*, 6(1), 41–80. https://doi.org/10.1007/BF02223791 - Lin, K. et al. (2025). Sleep-time compute. arXiv:2504.13171. https://arxiv.org/abs/2504.13171 - Martin, L. (2026). The components of agent memory. Anthropic, 6 November 2026. https://www.anthropic.com/research/components-of-agent-memory - McClelland, J. L., McNaughton, B. L. & O’Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. *Psychological Review*, 102(3), 419–457. https://doi.org/10.1037/0033-295X.102.3.419 - Payne, J. D. et al. (2009). The role of sleep in false memory formation. *Neurobiology of Learning and Memory*, 92(3), 327–334. https://doi.org/10.1016/j.nlm.2009.03.007 - Richards, B. A. & Frankland, P. W. (2017). The persistence and transience of memory. *Neuron*, 94(6), 1071–1084. https://doi.org/10.1016/j.neuron.2017.04.037 - Schneirla, T. C. (1944). A unique case of circular milling in ants. *American Museum Novitates*, 1253, 1–26. https://hdl.handle.net/2246/3733 - Shaikh, O., Mozannar, H., Bansal, G., Fourney, A. & Horvitz, E. (2025). Navigating rifts in human-LLM grounding. *ACL 2025*, 20832–20847. https://doi.org/10.18653/v1/2025.acl-long.1016 - Tse, D. et al. (2007). Schemas and memory consolidation. *Science*, 316(5821), 76–82. https://doi.org/10.1126/science.1135935 - Turpin, M., Michael, J., Perez, E. & Bowman, S. R. (2023). Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. *NeurIPS 2023*. https://arxiv.org/abs/2305.04388 - Wang, B. et al. (2025). Unveiling privacy risks in LLM agent memory. *ACL 2025*, 25241–25260. https://doi.org/10.18653/v1/2025.acl-long.1227