{"path":"research/what-agent-memory-research-says-about-agent-surfaces.md","content":"# What agent-memory research says about the agent surfaces\n\n**Status (2026-10-09): a reading with four proposals, none ruled (ledger Arc AE).** Written after the\nfounder asked whether a session's trouble taking the whole corpus into account is a problem of memory\nfor coding agents, and added: *\"any system that successfully and efficiently augments agents with\ngraph memory will likely have strong parallels to bear on the best future architecture for Deliberus'\nexternal agentic surfaces, in the very least.\"*\n\n**Read depth.** Four helpers read papers and product pages from 2024 to 2026. Claude read none of the\npapers. It checked that the reports' headline numbers stand in the page texts the helpers saved, and\nnothing more. A figure marked *excerpt* comes from a search excerpt and was not checked at all. The\nreports are kept, word for word, with a working note on harness memory outside this repository.\n\n## In plain language\n\nPeople who build memory for AI agents have measured five things. An agent-facing Deliberus should\ntake them as given.\n\n1. **An agent that owns a memory mostly does not look in it.** In one small study a coding agent with\n   a filled memory store made no memory call in 114 turns, while the same items delivered by its\n   harness at set moments arrived every time, 138 of 138 (arXiv 2607.20972). In another, when a memory\n   system answered wrongly, the evidence was sitting in its store about ten times as often as it was\n   missing (arXiv 2606.17328). This project has its own instance: a session wrote proposals that met\n   nine things already ruled or built ([what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 14), and at\n   least five of them were lines in the instructions it had loaded. The note holding the ruling behind the\n   worst of them is cited by all four notes that carry the proposals, and the session had opened it 49\n   minutes before it first wrote the proposal that meets it.\n2. **A summary handed back without its source loses.** A reader given a digest it cannot open scored\n   0.1 where the full method scored 58 (arXiv 2512.24601). Summaries tied to the raw text and checked\n   against it scored 60.1, against 55.2 unchecked (arXiv 2606.10662).\n3. **Updating and contradiction are what memory systems do worst.** Where a later fact conflicts with\n   an earlier one, plain keyword search scored 56 and three memory systems 14 to 28. With two steps of\n   reasoning nothing passed 6 (arXiv 2507.05257). A system that compresses its history has been seen\n   overwriting facts (arXiv 2410.10813). One that rewrote its own memory in a single step shrank it\n   from 18,282 tokens to 122 and did worse than with no memory at all (arXiv 2510.04618).\n4. **For plain recall, well-kept text with search does as well as a graph.** A graph helps where a\n   question hops across named things: 9.5 points on one such set (arXiv 2502.14802). It costs what its\n   extraction drops. In the one independent comparison found, a graph memory scored 22 points under\n   plain retrieval because facts had never reached the graph (arXiv 2601.07978, *excerpt*).\n5. **Handing a reader everything loses to handing it the right part.** Accuracy fell 30 to 60 percent\n   when a whole history of about 115,000 tokens was read in place of the evidence alone\n   (arXiv 2410.10813).\n\n## What this says about our surfaces\n\n**A tool an agent may call is half a surface.** [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md)\ndesigns a read-only set of tools. By the first finding, an agent will mostly not call them at the\nmoment that matters. What has been measured to work is delivery by the agent's harness at set\nmoments: before it states a conclusion, before it writes. So the tool set wants a companion built for\nthat use. It takes the sentence about to be written and returns the few claims that stand against it\nor already cover it.\n\n**What Deliberus holds that a memory layer lacks is the third finding.** Recall is a commodity. What\nstands against what, what was withdrawn, and which two opposite claims are both still standing is what\nthe claim graph is built to hold, and what every memory system tested does worst. An agent surface\nleads with opposition and standing. Nearest neighbours come second.\n\n**Every returned item carries its source.** The claim's id, its own words, its state and date, and a\nway to open what it rests on. The rule that no surface emits a critical question without both\nendpoints is this rule already, for one kind of item.\n\n**Computed, never rewritten.** The record is added to, and every reading over it is computed again.\nThe refusal of summaries rewritten in place\n([store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md) § 4b) now has a measured\nfailure behind it.\n\n**Return a region and say what was left out.** The map view's ten boxes are the same instinct, for\npeople.\n\n## Where the reading is thin\n\n- No study found uses research notes or arguments as the thing remembered. The tests are chat\n  histories, logs of web agents, and code.\n- The study of delivery by the harness is one small experiment.\n- The comparison of a graph memory with plain retrieval rests on an excerpt.\n- None of this has been tried on Deliberus. The agent-side test named in\n  [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md) (the substrate run) has not been run.\n\n## Proposals (ledger Arc AE, unruled)\n\n| | Proposal |\n|---|---|\n| AE1 | An agent surface includes a check that the agent's harness calls at set moments, beside the tools an agent may call |\n| AE2 | An agent surface leads with what stands against a statement and what was withdrawn; resemblance comes second |\n| AE3 | Every item an agent surface returns carries its id, its own words, its state and date, and a way to open what it rests on |\n| AE4 | An agent surface returns a small region and names what kind of thing it left out |\n\n## Cross-references\n\n[Practices for sustaining interconnected research](practices-for-interconnected-research.md) compares\nresearchers' working methods and distinguishes finding, applying, and maintaining knowledge; its\nsuggested transfers to Deliberus are proposals, not rulings.\n\n[agents-as-a-consumer-class.md](agents-as-a-consumer-class.md) (the read-only tool set, the substrate\nrun, why shipping is gated on the strength layer) ·\n[store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md) § 4 (the graph of real\nthings, computed from the claims) · [the-harness-conviction.md](the-harness-conviction.md) (machine\nreasoning's quality is a function of harness and grounding) ·\n[what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 14 (this project's own instance of memory present\nand not used) · [steadying-an-outside-judge.md](steadying-an-outside-judge.md) (what an outside agent\nwould read the record for).\n"}