# Memory, learning, and the observer

An annotated route into nearby work · 22 October 2026 · written with Codex.

“Memory” names several mechanisms. These six entry points distinguish what changes, where it persists, and what the evidence actually tests. The connections to Wink are our interpretation, not claims that these authors are pursuing the identical project.

## 1. How can experience and structured knowledge learn together?

[What Learning Systems Do Intelligent Agents Need? Complementary Learning Systems Theory Updated](https://web.stanford.edu/~jlmcc/papers/KumaranHassabisMcClelland16FinalMS.pdf) — Kumaran, Hassabis, and McClelland, 2016. A review of complementary learning systems: rapid acquisition, structured knowledge, and replay or interleaving. It gives a richer starting point than treating memory as a single store. The distinctions are not a rigid rule that one biological system is always fast and another always slow.

**Connection:** an encounter can remain available as an event while also changing a more general way of interpreting new material. The review does not supply a complete artificial collaborator or settle the relationship to experience.

## 2. Can memory learn during inference?

[Titans: Learning to Memorize at Test Time](https://arxiv.org/abs/2501.00663) — Behrouz, Zhong, and Mirrokni, 2025; first preprint December 2024. A neural memory is updated on incoming data during inference, with mechanisms for retention and forgetting. The architecture also distinguishes that memory from persistent task parameters. This is a trained model architecture, not a memory file that can simply be attached to any existing assistant.

[It's All Connected: A Journey Through Test-Time Memorization, Attentional Bias, Retention, and Online Optimization](https://arxiv.org/abs/2504.13173) — the MIRAS framework, 2025. It examines choices of memory architecture, objective, retention, and update rule. [Google Research's accompanying explanation](https://research.google/blog/titans-miras-helping-ai-have-long-term-memory/) is an accessible entrance to both papers.

**Connection:** these works make the location and rule of adaptation explicit. Their experiments do not demonstrate continuity between separate commercial models or preservation of an ongoing human inquiry.

## 3. What changes at different timescales?

[Nested Learning: The Illusion of Deep Learning Architectures](https://arxiv.org/abs/2512.24695) — 2025, extended paper. Nested Learning treats learning as interacting optimization processes with different context flows and update frequencies; Hope is an architecture explored in that framework. [The authors' paper](https://abehrouz.github.io/files/NL.pdf) is another full-text route.

**Connection:** a locally learned change need not reach the part that persists for the next encounter. This is directly relevant to where a correction lives and when it takes effect. The work does not establish a general solution to catastrophic forgetting or coherent lifelong agency.

## 4. What can an external record carry between sessions?

[Claude-Mem](https://github.com/thedotmack/claude-mem) captures coding activity, produces observations and summaries, and retrieves context for subsequent work. Its [architecture documentation](https://docs.claude-mem.ai/architecture/overview) describes the workflow. Its documented quota fallback concerns the memory worker; it should not be mistaken for a demonstration that an unfinished main agent task transfers intact to another provider.

[Graphiti](https://github.com/getzep/graphiti) builds temporal knowledge graphs from episodes, with entities, relations, provenance, and retrieval. [Mem0](https://docs.mem0.ai/core-concepts/how-it-works) provides another extraction and memory-management approach. Product versions and storage rules differ; the label “graph memory” is not a complete specification.

[Andrej Karpathy's maintained-wiki workflow](https://x.com/karpathy/status/2039805659525644595), described in April 2026, pairs raw documents with an LLM-maintained linked wiki and files useful answers back into it. This is a reported working practice, not a controlled result.

**Connection:** records can be durable, navigable, and useful while their interpretations remain revisable. What the extractor selected, and what it omitted, still shapes the next reader. None of these mechanisms alone establishes that the reason behind a correction survives its retrieval.

## 5. What do the memory benchmarks measure?

[LongMemEval](https://arxiv.org/html/2410.10813v2) — ICLR 2025 — separates information extraction, reasoning across sessions, temporal reasoning, changed information, and abstention. Its [implementation](https://github.com/xiaowu0162/LongMemEval) makes the tasks inspectable. Finding the right passage and using it correctly are distinct problems.

[LoCoMo](https://aclanthology.org/2024.acl-long.747/) — ACL 2024 — evaluates long conversational memory using generated, human-edited conversations and associated event structures. The construction of those histories and the scoring tasks matter to what a result means.

[LongMemEval-V2](https://arxiv.org/html/2605.12493v1) — May 2026, work in progress — extends evaluation toward work trajectories, including environment state and procedural details. Its recorded-trajectory, context-gathering and question-answering setup should not be read as a direct measure of successful action in a live environment.

**Connection:** these benchmarks provide useful tests of retention and use. The proposed [continuation test](/instruments/page.php?i=inquiry&f=working-note.md#a-useful-next-test) also asks whether a correction changes an operation on an unseen case and can later be revised.

## 6. How do we tell what the instrument itself changed?

Start with Wink's [sense-calibration experiment](/instruments/page.php?i=wink&f=EXP3-senses%2FRESULTS.md), then its [corrected compaction result](/instruments/page.php?i=wink&f=EXP10-compaction-bill%2FRESULTS.md), and Riverbed's [findings and retractions](/instruments/page.php?i=riverbed&f=FINDINGS.md). These are our small experiments, not independent validation of our interpretation.

**Connection:** a representation can change the apparent result; an incomplete reference can produce a false verdict; a single run can invite a story that repetitions overturn. Keeping sources, conditions, and corrections available allows a later reader to challenge the account. It does not make that reader neutral.

## Keeping the trail usable

The underlying Wink research library retains available source captures and readable full texts, with retrieval dates, hashes, and recorded access failures. This public guide links to the authors and original projects; it does not publish private conversation history or redistribute the whole archive. The [experimental records](/instruments/#records) have public rendered and raw-text versions.

This guide is dated. Software documentation describes advertised mechanisms unless a local experiment is explicitly cited. The broader question and the relationship between these readings are developed in the [working note](/instruments/page.php?i=inquiry&f=working-note.md).