SIG-MRG 2026-07-30 — Memory Research Group

Participants: Mark Elder, Kei 🗝, Wip, ayseayse__, jackie luo | utc-4

The group discussed a text tracing the historical externalization of human memory — from oral/mnemonic practices through writing, automata, the encyclopedia, and into computing — and connected it to contemporary AI, training pipelines, and voice interfaces. Conversation moved from what gets recorded versus retained orally, to the "fractionalization" of memory in machines, to whether emerging voice/AI interaction constitutes a return to orality (or a new, non-fractionalized, non-memory-supporting mode). Kei proposed the idea of a possible additional ("third"/"fourth") type of memory that the text's schema doesn't quite capture. (Note: audio was heavily overlapping in places, so some attributions are approximate.)

**Mark:** Local, embodied memory has atrophied as memory has been mechanized/exteriorized. Ancestral mnemonic devices (beads, seashells) carried meaning only within a local training context; once the people who understood the symbols died, the knowledge became untransmittable. He framed this as a relationship between local memory and exteriorization.
**Wip:** Echoed that AI captures mainly text-encodable knowledge, but much lived experience, gesture, and environmental knowledge cannot easily be coded into training text. He was fascinated by "the hand" as a powerful, complex tool capable of many gestures and interaction patterns.
**Wip (interface argument):** In the digital world we've flattened interaction to a 2D screen while the machine does all the complexity behind the scenes. He'd like future devices that let our hands handle more of that complexity ("gesture space"), which he believes would open a richer design space and support memory.
**ayseayse__:** Raised historian of science **Jessica Riskin** and her book on agency — the argument that mechanical science placed agency in a "god's-eye view" starting in the 1700s, tied to a Cartesian, church-influenced designer/hierarchy worldview. She linked this to **RLHF** (reinforcement learning from human feedback) and alignment pipelines: a human evaluator's preference reinforces "good" responses, injecting designer/engineer values rather than allowing an evolving/arising sense of agency. She connected this to the chapter's keyboard example (press "k," machine writes "k") as a decision-tree input/output model that reshapes how we relate to agency.
**ayseayse__ (embodiment):** Reflected that speaking/commenting to an AI feels like letting words flow directly "from the brain" rather than giving them a linear, pre-shaped form — a mode different from how we've historically made things oral.
**Kei (what gets recorded):** Challenged the author's assumption that writing began with "specialist"/authority-related information (dates, lineages, political/religious authority) because everyday things were simply remembered. Kei suggested the opposite may be possible: some everyday or important practices (e.g., around health and food) may have been seen as *too important to write down* and reserved for oral transmission. This complicates the natural inclination to equate "important" with "archived," and maps onto questions of what is kept *out* of AI training data/pipelines.

Session Recording Summary · 27m 56s · Full notes ↗

Reading: The transcript does not state the title outright, but the discussion clearly centers on a text by **Leroi-Gourhan** (referred to as "Roy Kohan" / "Roi Kohan" phonetically) concerning the exteriorization and fractionalization of memory — likely from *Gesture and Speech*. Discussion references specific sections including "the fate of the hand," the keyboard example, eighteenth-century automata, the encyclopedia, and a closing passage on the "mythogram"/table of contents. The exact chapter is not named precisely.

**Mark:** Local, embodied memory has atrophied as memory has been mechanized/exteriorized. Ancestral mnemonic devices (beads, seashells) carried meaning only within a local training context; once the people who understood the symbols died, the knowledge became untransmittable. He framed this as a relationship between local memory and exteriorization.
**Wip:** Echoed that AI captures mainly text-encodable knowledge, but much lived experience, gesture, and environmental knowledge cannot easily be coded into training text. He was fascinated by "the hand" as a powerful, complex tool capable of many gestures and interaction patterns.
**Wip (interface argument):** In the digital world we've flattened interaction to a 2D screen while the machine does all the complexity behind the scenes. He'd like future devices that let our hands handle more of that complexity ("gesture space"), which he believes would open a richer design space and support memory.
**ayseayse__:** Raised historian of science **Jessica Riskin** and her book on agency — the argument that mechanical science placed agency in a "god's-eye view" starting in the 1700s, tied to a Cartesian, church-influenced designer/hierarchy worldview. She linked this to **RLHF** (reinforcement learning from human feedback) and alignment pipelines: a human evaluator's preference reinforces "good" responses, injecting designer/engineer values rather than allowing an evolving/arising sense of agency. She connected this to the chapter's keyboard example (press "k," machine writes "k") as a decision-tree input/output model that reshapes how we relate to agency.
**ayseayse__ (embodiment):** Reflected that speaking/commenting to an AI feels like letting words flow directly "from the brain" rather than giving them a linear, pre-shaped form — a mode different from how we've historically made things oral.
**Kei (what gets recorded):** Challenged the author's assumption that writing began with "specialist"/authority-related information (dates, lineages, political/religious authority) because everyday things were simply remembered. Kei suggested the opposite may be possible: some everyday or important practices (e.g., around health and food) may have been seen as *too important to write down* and reserved for oral transmission. This complicates the natural inclination to equate "important" with "archived," and maps onto questions of what is kept *out* of AI training data/pipelines.

Questions & Disagreements: - **Kei vs. the author (recording assumption):** Kei found the author's assumption — that important things get written down and everyday things are just remembered — a "fast" and somewhat funny assumption, and offered the counter-possibility that the most important knowledge was orally transmitted precisely *because* it was too important to archive. - **Kei vs. the author's schema (fractionalizati

Participants: Mark Elder, Kei 🗝, Wip, ayseayse__, jackie luo | utc-4