Where AI Memory Lives

The recall system

Espalier-harness, my workflow harness for Claude Code, keeps a project’s judgment in committed files: a short memory index, a folder of lessons, and a catalog of sharp edges, the mistakes a project has already paid for once. A hook brings the relevant lessons back into a session when it detects a trigger, and the /recall command pulls them when you ask. Recall returns up to four candidates from two rankers, and you choose. It does not have to wait to be asked, either: recall can be automated or scheduled in workflows, written into an agent’s instructions, added to other skills and commands as a task, or prompted by a note in the docs every session reads.

Two lessons came out of building it. First, a hit is the nearest match, not proof that the topic is covered: ask an off topic question and the search still returns something that looks confident. Second, memory is found by its hazard, not its task. A lesson is filed under what went wrong, so asking “what could go wrong here?” finds it where “how do I do this?” does not. In one small test, questions worded as the task found none of the four lessons that applied; worded as the hazard, they found all four. The design is laid out in the repository’s memory systems guide.

The research knowledge base

The second system indexes what my research projects have learned: their findings, decisions, plans and memory notes, in one SQLite database. Nothing is summarized by a model. Every entry is the original text, with the file and line it came from. Search is plain keyword search, but the entries are linked: when a later finding supersedes or refutes an earlier one, the knowledge base records it, so an answer arrives with its standing. A memory note shows whether the findings it leans on still stand, so a retraction reaches the notes that depend on it.

It also taught me something about asking. The same question written five ways (as the task, the hazard, plain words, the field’s own terms, and the title you expect to find) puts the right answer near the top more often than the question alone. Scored on 139 questions that crossed projects, five phrasings put the right answer in the top five for 96, and the question alone for 80. How big the gain is depends on whose answer labels you trust, but it was positive under every set of labels: 6 to 16 more questions.

Memory in compute

Both systems keep memory as text, outside the model. That works, but every recall is a search, and every search can miss. I think the future may store memory in calculation and compute, not in memory space. Weights are storage too, so the shift is not away from storage altogether. It is from keeping text to read again toward writing what was read into the model itself.

Research is already moving that way along a few lines.

The model learns while it reads. Titans (Google, 2024) adds a neural memory whose weights keep updating as the model reads, with attention serving as short-term memory beside it. End-to-End Test-Time Training for Long Context (2025) has a model keep training on the text in front of it, so the text ends up in its weights. At 3 billion parameters it kept pace with full attention as the context grew, with the same speed at any length.

A network writes the weights. Doc-to-LoRA (Sakana AI, 2026) turns a document into a small set of adapter weights in one pass, so later questions no longer reread the document. The boldest name so far is Infinite-Parameter LLMs (September 2026), which generates weight adjustments from live data during a session. “Infinite” there means the range of weights it can produce, not unlimited storage: the model stays a fixed size, and the authors call their results preliminary.

Storage gets cheap instead. Not everything points one way. Memory Layers at Scale (Meta, 2024) adds a trainable lookup table of up to 128 billion parameters at almost no extra compute, with the largest gains on factual tasks. That is more memory space, not less.

All of this is research scale: models of a few billion parameters, with results mostly reported by their own authors.