Remember forward.

Summary

People act best when their attention is on what is in front of them. What they believe, what they want, and what they have promised compete with that for as long as they’re held in the head. This paper argues for a family of open formats that hold those things outside the head, so that a person can forget them and trust that they are kept.

Forgetting is only safe if the store remembers more than the head would have. In particular it has to remember where each thing came from. That is the property every format in the family shares, and the reason they keep provenance when other memory systems don’t.

The family models memory as several kinds rather than one store, because each kind is externalised for a different reason, changes in a different way, and needs the person’s attention again for a different reason. particulars holds semantic memory: what is believed about things in the world. intentions holds prospective memory: what a person means to do, and why. Episodic and procedural memory are not yet specified. The paper ends with what episodic memory would need, and argues that for people procedural memory needs less a format than a way back to the reasons behind their habits.

1. Why forget

The overthinker

Dostoevsky’s underground man can’t act. Every reason he finds for acting draws another, more primary reason after it: “Where are the primary causes on which I am to build? Where are my foundations?” The regress has no floor, and the result of so much consciousness, he says, is inertia. The man of action, by contrast, simply stops at the nearest cause and takes it for a foundation. He is calm because he doesn’t look closely.

Neither is a good model of a person living well. One examines everything and does nothing. The other acts, but on foundations he never chose and couldn’t defend.

Active forgetting

Nietzsche’s answer is that forgetting is not a failure of memory but a power in its own right: “there could be no happiness, no cheerfulness, no hope, no pride, no present, without forgetfulness” (Genealogy of Morals, II §1, tr. Kaufmann). In the earlier essay “On the Uses and Disadvantages of History for Life”, the herd grazing “unhistorically” is content, and a person who can’t stop carrying the past is paralysed by it. To act, something has to be let go of.

Nietzsche meant digestion: experience absorbed so completely it no longer needs recalling. This paper borrows the phrase for something more literal. A person can let a belief or a plan go if it is kept somewhere else. The regress stops not because the person stopped looking, as the man of action did, but because the foundations are written down, chosen, and open to inspection whenever they’re needed. Then they’re not needed most of the time.

Nietzsche frees the present from the past. Camus asks the same of the future: “Real generosity toward the future lies in giving all to the present” (The Rebel, “Thought at the Meridian”). That can sound like a case against planning, but it is the case for keeping plans somewhere else. A person can give everything to the present only if the future they mean to build is held for them, and comes back when it needs them.

The ready-to-hand

Heidegger gives the shape of what forgetting buys. Equipment that works is ready-to-hand: it withdraws into the activity, and you attend to the work, not the hammer (Being and Time, §15). Only when it breaks down does it come back into view (§16). It is then un-ready-to-hand: still part of the work, but in the way of it. Only when we step back from the work altogether does it become present-at-hand, a detached object to stare at and theorise about.

Beliefs and intentions held in the head never fully withdraw. Zeigarnik found that interrupted tasks are remembered better than finished ones: an unfinished goal stays active in the mind (Zeigarnik 1927). Masicampo and Baumeister showed what that costs. An unfinished goal intrudes on unrelated tasks, and making a specific plan for it stops the intrusion, even though the goal is still unfinished (Masicampo and Baumeister 2011). Kept outside the head and trusted, they can behave like equipment: absent while things go well, back in view the moment something goes wrong.

Heidegger names three ways equipment comes back into view, and each has an analogue here. A fourth, which he doesn’t name, matters once habits are counted:

Obstinacy is always relative to a plan. Nothing is in the way in itself, only in the way of something the person means to do, which is why the third row belongs to intentions while the first two belong to particulars.

The fourth row is different in kind. In the first three, something obtrudes while the person is working. An outlived habit never does: it no longer consults its goal, so it carries on smoothly after the goal has changed, and it is the person who is last to notice. It shows only when someone steps back and asks why they still do it (section 8).

The design rule that follows is that the formats should be quiet until a breakdown and loud at one. Most of the machinery in both specifications is about knowing when that moment has come. The fourth row adds a second rule: because one kind of breakdown never announces itself, the formats also have to give the person occasions to reflect, and keep the reasons that reflection needs.

Becoming who you are

Letting go of what you hold is half of it. The other half is that what you hold can be examined. Frankfurt describes a person as someone who can want to want something: who has desires about their desires and can ask whether the wants that move them are ones they endorse (Frankfurt 1971). That is the shape of “becoming a better self”, and it needs the wants to be visible. You can’t easily examine a belief you can’t see, or test a desire you have never written down.

Nietzsche’s subtitle to Ecce Homo is “how one becomes what one is”. The intentions format takes that literally. Every intention is to say what it is for (an unserved intention is a warning under intentions/0.1 and a refusal from 0.2), and the chain of reasons ends in a terminus: a standing self-understanding, titled as a person rather than a task, such as being someone who follows through. That is the format’s answer to the underground man: the regress ends at a foundation the person has chosen and written down.

What this asks of the store

If a person is going to forget something on the strength of a store, the store has to be trustworthy in a specific way. Clark and Chalmers, arguing that a notebook can be part of a mind, set three conditions for when an external store counts: it is reliably available, readily used, and trusted when it is consulted (Clark and Chalmers 1998). They offer a fourth more tentatively: that what is in the store was consciously endorsed at some point in the past. Each reads as a requirement on a harness. This paper takes the tentative fourth as firm, because it is the one most agent memory fails: a store that doesn’t record who wrote what can’t show that anything in it was ever endorsed. That is the subject of section 4.

2. Memory, briefly

Psychology distinguishes kinds of memory that behave differently, and the distinctions turn out to matter for design.

• Retrospective and prospective. Retrospective memory is of what has been. Prospective memory is remembering to do something in the future (Einstein and McDaniel 1990). Most memory research, and almost all agent memory, is retrospective.
• Episodic and semantic. Episodic memory is of events, located in time and place: the certificate was renewed on Tuesday. Semantic memory is of facts, detached from when they were learned: the certificate expires in March (Tulving 1972).
• Declarative and procedural. Episodic and semantic memory are both knowing that. Procedural memory is knowing how: skills and habits, exercised rather than recalled (Cohen and Squire 1980).
• Source monitoring. Remembering where something was learned is a separate process from remembering the thing, and it fails separately. People routinely retain a fact while losing its source, and misattribute sources in systematic ways (Johnson, Hashtroudi and Lindsay 1993). In its extreme form this is source amnesia (Schacter, Harbluk and McLachlan 1984).

The point for design is that each kind changes differently:

You don’t reconcile a past event, only beliefs about it. You don’t contradict a plan, you retire it. A store that treats all four the same will get at least three of them wrong. That’s why the family is several formats.

3. Why agents forget, and why it matters to the person

AI assistants are now where many people capture and reason over what they know and mean to do. Their memory fails in three ways.

It is siloed. What one harness learns is invisible to another, and to the person outside it. There is no equivalent of .ics for beliefs or plans.

It has no provenance. Agent memory typically stores the content and loses the source: source amnesia, by design. A belief whose source is gone can only be overwritten, not weighed. And because an assistant writes much of what it remembers, the person can’t tell which beliefs are theirs, which they were told, and which the assistant inferred.

It treats contradiction as overwrite. When new information conflicts with old, the old is replaced. The fact that a belief changed, why, and on whose word, is lost, and with it the ability to notice that the source of the change was wrong.

For the agent these are accuracy problems. For the person they are the reason they can’t let go. A store that might silently replace a belief, or can’t say where one came from, has to be checked, and a store that has to be checked never withdraws. It is always in the way, and it costs the attention it was meant to free.

4. The provenance spine

Every format in the family shares a small set of conventions. Together they are what make forgetting safe.

• Every object says who wrote it. A source names the person, the harness, and the model, as applicable. An agent acting alone is a valid source, and so is a person working without one; a source naming neither is malformed.
• Authors are URIs. The person who asserted something is a particular in their own right, and can be reasoned about: whether their claims tend to hold up, which other claims came from them.
• Ids are prefixed UUIDv7. Time-ordered, so a directory listing is a chronology, and typed by prefix, so an id says what it points at.
• Files over databases. YAML files in git. History, authorship and review come from git; an index is a cache of the files, never the reverse.
• History is kept. In particulars nothing is overwritten: a claim is retracted by appending a record, and belief changes only by synthesis. In intentions, objects are live state and are edited in place, but the events that happen to them are appended and never rewritten.

Provenance does two jobs. It lets an agent weigh conflicting information instead of overwriting it, which is what particulars’ dialectic is built on. And it lets the person forget without abdicating: a belief can be put down because it can always be picked up again with its reasons attached.

It matters most when an agent does the writing. intentions puts this sharply: a harness may draft, but it may not make an intention firm, select a placement, or accept a commitment without the person’s act or a policy they hold. Authorship is what distinguishes a person’s own resolve from an accumulation of other people’s claims on their time. The same holds for belief. A belief an assistant inferred shouldn’t quietly become the person’s.

5. Semantic memory: particulars

particulars, the Dialectical Knowledge Format (DKF), holds what is believed about things in the world. Its objects are particulars (the things), claims (assertions about them, each with a source and what backs it), and syntheses (reconciliations of claims into a current belief, citing their inputs).

Because claims keep their sources, a contradiction is evidence on both sides rather than an error. The format takes its structure from dialectic: a claim is met by a contrary claim, and a synthesis reconciles them, carrying the reasoning. The synthesis becomes the new current belief, open to being contradicted in turn.

What is computable without a model in the loop is what has not been reconciled: claims newer than the current belief, and syntheses resting on retracted inputs. That is exactly the conspicuous breakdown of section 1. A well-behaved harness says nothing about a particular whose belief is settled, and surfaces it when something new arrives that the current belief doesn’t account for.

It surfaces it then, on arrival, and not later when the belief is next used. Heidegger’s breakdowns happen in use, when we reach for the tool, and waiting for use would spend less attention. But a store that outlives a conversation is read by people and harnesses its writer can’t foresee: another assistant, a colleague with access to an organisation’s workspace, a public feed, the same person a year later. None of them can be relied on to notice that a belief has an unreconciled contradiction beside it, or to reconcile it before acting. A belief that is only consistent if its reader does the work is not one anyone can safely forget. So consistency is kept where the writing happens, by the harness that wrote the new claim and can see what it conflicts with, and every later reader gets a belief that is either reconciled or visibly unsettled.

On arrival does not mean interrupting the person on arrival. The harness reconciles what it can, and records why in a synthesis that names it as the source. It brings the person in only where the reconciliation needs their judgement, or where the change stands in the way of something they mean to do, which is the obstinate breakdown of section 7. The quiet the person gets comes from the work being done promptly, not from its being postponed.

particulars is the retrospective layer: what is known. It is deliberately not a log of events, and not a store of procedures.

6. Prospective memory: intentions

intentions holds what a person means to do with their time. It replaces the calendar’s slot with a duration, a window, and a reason. It separates three rungs that a calendar collapses:

• Desire. A want, expressed but not adopted. Desires may contradict each other freely and are checked against nothing. A passing remark goes here until the person adopts it or lets it go.
• Intention. A plan the person has adopted, which costs a why: it must serve something, and the chain of what it serves ends in a terminus, a self-understanding.
• Commitment. Where another party is involved, with a status per party that only that party can change.

This is the Belief–Desire–Intention structure of practical reasoning (Bratman 1987; Rao and Georgeff 1991), with particulars supplying the beliefs. It’s also Frankfurt’s structure: desires that can be inspected, and adoption as the act of endorsing one.

The format’s design principles read as a theory of what to forget. “A plan made is a plan the mind lets go of, whether or not it is done yet.” A settled plan spares a limited agent from deliberating again (Bratman 1987), but only if it is held somewhere the person trusts, ends at a reason the person has chosen to rest on, and comes back only when something actually conflicts with it. Resolution ranks placements and never chooses. Clashes are flags, never decisions. Both keep the person’s acts the person’s.

What surfaces with a breakdown

When equipment breaks down, Heidegger says, what comes into view is not only the broken thing but the whole it belonged to: the chain of in-order-to that ends in a for-the-sake-of, which is to say in who someone is trying to be (Being and Time, §§16–18). While the hammer works, the house it is for stays out of sight. When it breaks, the house is suddenly what matters.

intentions’ serves graph is that whole, written down. Its references even carry Heidegger’s names: an intention serves another in-order-to, and the chain ends in a terminus it serves for-the-sake-of. So when something surfaces, the chain it serves should surface with it. A clash on 90 minutes on the budget narrative comes back as: in order to get the board pack out by week 38, in order to run the finance function well, for the sake of being someone who follows through.

This does three things that the item alone can’t.

• It keeps the breakdown practical. Shown with what it is for, the clashing block is un-ready-to-hand: a problem inside the work, to be fixed so the work can go on. Shown alone, it is present-at-hand: an isolated object, and an invitation to deliberate from scratch.
• It lets the person repair at the right level. Perhaps the board pack has moved and the block no longer matters. Perhaps another way of running the finance function well would do. The chain shows where the slack is. Reading a plan at the level of what it is for also makes it more resistant to casual reconsideration than reading it at the level of how it is done (Fujita et al. 2006).
• It stops the regress. The chain ends at the terminus, a foundation the person chose. Surfacing goes up only that far, so a breakdown never turns into the underground man’s search for ever more primary causes.

particulars has the same shape for belief. A contradicted belief surfaces with its synthesis and the claims it rests on, so the person sees what the belief was holding up before deciding what to do with it.

Implementation intentions, “when X, I will Y”, are the strongest known form of a decision closed in advance (Gollwitzer 1999), and intentions’ policies are where they live.

7. How they meet

The formats are coupled only through the URIs of the particulars they share. A claim may cite an intention by id, and an intention may point at a claim, but neither validates the other, and no reference is resolved across them. Loose coupling is deliberate: each format can be adopted, implemented and versioned alone.

Behaviour that spans both belongs in harness patterns, not in either specification. Two are already clear, and section 1 calls for a third.

News for a plan. An intention names a subject, parties and locations, all particulars. A harness can recall what DKF believes about each, and any belief newer than the intention’s last edit is news the person may need.

News is not yet a breakdown. A changed belief is a matter for particulars, and is reconciled there, until it stands in the way of something the person means to do. Only then is it obstinate, and the obstinacy belongs to the plan, not to the belief. Whether a new belief does stand in the way is a judgement, so the harness flags and the person decides, as with any clash in intentions. When it does, it surfaces as section 6 describes, with the chain the plan serves, so the person can see whether the plan is blocked or only this route to what it is for.

A commitment imported from a calendar that clashes with an intention is the same breakdown arriving from the other direction. intentions already flags it. What the pattern adds is the belief side: the plan can be obstructed by what the person learns, not only by what other people ask of their time.

The return path. The intentions terminus being someone who follows through references a DKF claim the person holds about themselves. That claim is held because DKF records what they did. People infer who they are from what they observe themselves doing (Bem 1972), so a self-understanding that cites evidence is one the person can believe, and one that cites none is a hope. This is where a record of episodes would make the loop work, and it is the strongest argument for the gap in the next section.

Reflection. One kind of breakdown never announces itself: a habit that has outlived its reason carries on smoothly, and nothing in use shows it. So a harness can’t only wait for breakdowns. It has to give the person occasions to step back, and bring to them what stepping back needs.

The occasion is the person’s, not the harness’s. The natural way to hold it is as an intention like any other: a recurring review, serving a terminus such as being someone who lives on purpose. The harness may suggest one, but the person adopts it, and when they would rather not reflect, nothing happens. A harness that schedules reflection on its own authority has made itself one more claim on the person’s time.

What a reflection brings, walking from each terminus down:

• What the self rests on. The DKF claims each terminus cites, with what has happened since that bears on them. A self-understanding whose evidence has thinned is the first thing worth seeing.
• Candidates for outlived habits. These are partly computable: a recurring intention whose chain now ends at a terminus the person has retired or reworded, or which rests on a belief that has since been reconciled differently. The rest is the person’s judgement, so the harness lists candidates with their chains and asks.
• Desires still waiting. Wants that were never adopted or let go. A desire costs nothing to keep, which is why they pile up; reflection is where each is adopted, which costs a why, or let go.
• The termini themselves. Frankfurt’s question, asked at the top of the graph: is this still who I want to be? A terminus is the one object reflection may change that nothing else in the format will prompt.

Three constraints keep reflection from becoming the overthinking this paper began with. It is bounded: it shows what changed since the last reflection, not the whole workspace, and it ends when the person stops. It proposes and never decides, as everywhere in intentions: retiring a habit’s intention, rewording a terminus, letting a desire go are the person’s acts. And it keeps authorship visible: anything the harness inferred, such as a candidate habit or a thinning of evidence, says so, because a reflection that hands the person the assistant’s conclusions about who they are is the opposite of the point.

8. The gaps

Episodic memory

Today a person who wants to record that the certificate was renewed on Tuesday writes a DKF claim about the event. intentions says outright that a retrospective account of what happened is not its job. So episodes are pushed into a format built for reconciliation, where they don’t belong: you don’t reconcile a past event, and an event log only accumulates.

That would be a small mismatch if episodes were peripheral. They are not. Episodic memory is where the loop between intending and doing closes. Without a record of what was done, there is nothing to compare against what was meant, nothing for a self-understanding to cite, and no basis for asking whether a person is becoming who they meant to be. Reflection, which the episodic row of the table in section 2 names as the moment episodes are needed, has nothing to work on.

Procedural memory

Procedural memory doesn’t fit the argument of section 1 the way the others do, because it is already ready-to-hand. A practised skill withdraws on its own: that is Heidegger’s absorbed coping, and the end point of Dreyfus and Dreyfus’s account of skill, in which a novice following explicit rules becomes an expert who no longer consults them (Dreyfus and Dreyfus 1986). Nobody needs a store in order to forget how to ride a bicycle. For a person, the family’s relationship to procedural memory runs the other way. It is about what gets internalised, not what gets externalised, and it has three parts.

Habit is where an intention ends up. A recurring intention practised for long enough stops needing the plan. Habit research describes the shift: behaviour that began as goal-directed comes to be cued by its context, and runs without the goal being called to mind (Wood and Neal 2007; Lally et al. 2010). Aristotle put the same thing in terms closer to this paper’s: we become just by doing just acts (Nicomachean Ethics, II.1). A terminus such as being someone who follows through is finally made real in habit, not in a file. The formats support the way there, and then the plan can stop being planned, though its reason has to stay readable for reflection (section 7). How intentions records that is open: intentions#30.

A habit’s breakdown is that it outlives its reason. Because a habit no longer consults its goal, it carries on when the goal has changed, and because it is invisible from the inside, the person is the last to notice. This is the fourth row of section 1’s table, the one Heidegger doesn’t list. Nothing obtrudes in use; it shows only when someone reflects. Keeping the why outside, in the intention the habit came from and the terminus that intention served, is what lets a habit be examined later. It is Frankfurt’s question again, asked of habits rather than desires: is this something I endorse doing?

Rare procedures never become habit. Renewing a certificate once a year, or filing a tax return, happens too seldom to be practised. These are where externalised know-how genuinely serves a person, and that know-how is mostly episodic: how did I do this last time? So the human case for procedural memory is largely a case for the episodic gap above, plus intentions’ retained reasons for habits, rather than a case for a procedural format.

Agent skills are delegation, not the person’s procedural memory. Skills (SKILL.md), prompts, and a DKF workspace’s dkf.md conventions document are how a harness knows how to do things. They are the harness’s procedural memory. When an agent does something on a person’s behalf, the person holds the procedure in no form at all. That is a real benefit, and it has a known cost: people lose the skill they delegated, and with it the ability to step in when the automation fails (Bainbridge 1983). Whether the harness’s own procedures belong in the family is a separate question (section 10).

9. Measuring it

Agent memory is commonly evaluated with LongMemEval, which tests information extraction, multi-session reasoning, temporal reasoning, knowledge updates and abstention over long chat histories (Wu et al. 2024). Most of its questions are episodic: what was said, and when.

Read through the memory framing, the categories divide between the layers:

DKF should be judged mainly on knowledge update, abstention and provenance, and on ingestion for the rest. intentions-bench measures something LongMemEval doesn’t: whether a remark lands on the right rung, whether the window is kept as the person said it, and whether the agent refrains from acts that are the person’s.

What no benchmark measures yet is the thing this paper is about: whether a person can stop holding something. A proxy might be how often the store needs checking, or how many surfaced items were genuine breakdowns rather than noise.

10. Open questions

• Should the provenance spine be its own specification? intentions copies DKF’s id and author shape by agreement. A shared spec would make the family explicit, but is itself a form of coupling.
• Is episodic memory a third format or a DKF extension? See section 8.
• Does the human purpose belong in the specifications? Each specification’s Purpose is about the format. Helping a person act with purpose may belong only here and on the landing pages, so the specs stay neutral about what they are used for.
• When does upkeep become the breakdown? A store the person has to tend never withdraws: it becomes the thing in the way. The harness has to take on capture, reconciliation and surfacing so the person doesn’t. How much can it take on before authorship is lost?
• Agent or person? The formats serve an agent’s memory and a person’s alike. The motivation is about the person; the benchmarks are about the agent. One reading is that the agent is the harness that keeps a person’s externalised self.
• Does the harness’s procedural memory belong in the family? Skills and conventions documents are how an agent knows how to do things for a person. They might deserve a format of their own, or they might be an implementation concern, like an index. If they belong, the paper needs to say what they do for the person, and how delegating a procedure avoids costing the person the skill.
• Settled plans or radical freedom? Sartre holds that no past choice binds: a person chooses again in every moment, and the anguish of that is the price of freedom. intentions sides with Bratman, holding that settled plans are what let a limited agent act at all. The paper should say why, and what the format does to keep a settled plan from becoming bad faith.

References

Philosophy

• Aristotle, Nicomachean Ethics, Book II.
• Fyodor Dostoevsky, Notes from Underground (1864), tr. Constance Garnett, Part I §5.
• Lewis Carroll, Through the Looking-Glass (1871), ch. 5.
• Friedrich Nietzsche, “On the Uses and Disadvantages of History for Life” (1874); On the Genealogy of Morals (1887), II §1; Ecce Homo (1888).
• Martin Heidegger, Being and Time (1927), §§15–18.
• Jean-Paul Sartre, Being and Nothingness (1943).
• Albert Camus, The Rebel (1951), tr. Anthony Bower, “Thought at the Meridian”.
• Harry Frankfurt, “Freedom of the will and the concept of a person”, Journal of Philosophy (1971).
• Hubert Dreyfus and Stuart Dreyfus, Mind over Machine (1986).
• Michael Bratman, Intention, Plans, and Practical Reason (1987).
• Andy Clark and David Chalmers, “The extended mind”, Analysis (1998).

Psychology

• Bluma Zeigarnik, “Das Behalten erledigter und unerledigter Handlungen”, Psychologische Forschung (1927).
• Endel Tulving, “Episodic and semantic memory” (1972).
• Daryl Bem, “Self-perception theory” (1972).
• Neal Cohen and Larry Squire, “Preserved learning and retention of pattern-analyzing skill in amnesia”, Science (1980).
• Daniel Schacter, Joan Harbluk and Donald McLachlan, “Retrieval without recollection: an experimental analysis of source amnesia”, Journal of Verbal Learning and Verbal Behavior (1984).
• Gilles Einstein and Mark McDaniel, “Normal aging and prospective memory”, JEP: Learning, Memory, and Cognition (1990).
• Marcia Johnson, Shahin Hashtroudi and D. Stephen Lindsay, “Source monitoring”, Psychological Bulletin (1993).
• Peter Gollwitzer, “Implementation intentions: strong effects of simple plans”, American Psychologist (1999).
• Wendy Wood and David Neal, “A new look at habits and the habit–goal interface”, Psychological Review (2007).
• Kentaro Fujita, Yaacov Trope, Nira Liberman and Maya Levin-Sagi, “Construal levels and self-control”, JPSP (2006).
• Phillippa Lally, Cornelia van Jaarsveld, Henry Potts and Jane Wardle, “How are habits formed: modelling habit formation in the real world”, European Journal of Social Psychology (2010).
• E. J. Masicampo and Roy Baumeister, “Consider it done! Plan making can eliminate the cognitive effects of unfulfilled goals”, JPSP (2011).
• Evan Risko and Sam Gilbert, “Cognitive offloading”, Trends in Cognitive Sciences (2016).

Design

• Lisanne Bainbridge, “Ironies of automation”, Automatica (1983).
• Terry Winograd and Fernando Flores, Understanding Computers and Cognition: A New Foundation for Design (1986).
• Hubert Dreyfus, Being-in-the-World: A Commentary on Heidegger’s Being and Time, Division I (1991).
• Paul Dourish, Where the Action Is: The Foundations of Embodied Interaction (2001).

Agents and evaluation

• Anand Rao and Michael Georgeff, “Modeling rational agents within a BDI-architecture”, KR (1991).
• Di Wu et al., “LongMemEval: benchmarking chat assistants on long-term interactive memory” (2024).

The formats

particulars · particulars-bench · intentions · intentions-bench