Learn · AI methods

Build a research OS that stays alive: agent memory, and who prunes it

The last decade's advice was to build a second brain: a vault of linked notes that thinks with you. The new advice is to hand that vault to an AI agent as its memory. This guide covers why research vaults die, how the four main agent-memory designs work and where each one leaks, and a way to run yours like a garden: the agent proposes, you prune, and the record shows who decided.

Why research vaults become graveyards

Most research vaults do not fail loudly. They fail the way a garden fails: quietly, from neglect that looked like growth. The patterns repeat across the Obsidian and Zettelkasten communities so consistently that they have names.

• The collector's fallacy. Saving a paper feels like progress, but as Christian Tietze put it on zettelkasten.de, knowing about something is not the same as knowing it. Collecting gives an instant reward; working through a source can take longer than reading it. So the vault fills with PDFs no one has processed.
• Orphan notes. One PhD student on the Zettelkasten forum reported that after six months, roughly nine out of ten notes in the vault had never been linked to anything. A note with no links is a note the system can never surface again.
• The notes-to-manuscript gap. A thesis writer's field report on zettelkasten.de landed on an uncomfortable truth: atomic notes make a useful vault, but they do not make chapter sections. The writing was still theirs to do. Notes support an argument; they do not assemble one.
• Maintenance becomes a second job. Folder trees twelve levels deep, workflows with five plugins, sync conflicts between devices, a Zotero-to-Obsidian pipeline that breaks when a plugin updates. Time spent tending the system is billed against time spent thinking with it, and the critics' question is fair: at what point is the second brain an excuse for not using the first one?

None of this means the method is broken. Other field reports describe theses finished on schedule precisely because the writer never started a chapter from zero. What separated those vaults from the graveyards was tending: a regular, deliberate practice of linking, revising, and discarding.

Four ways to give an agent memory

A language model starts every conversation blank. Everything it "remembers" between sessions lives in some external store that is loaded back in, and there are four main designs for that store. Each solves a real problem, and each leaks somewhere that matters for research. If the terms below are new, the AI-concepts map covers them from the ground up.

1 · Vector retrieval (RAG)

Your notes are split into chunks, each chunk is turned into an embedding (a list of numbers standing in for its meaning), and when you ask a question the system retrieves the chunks that sound most alike and hands them to the model. It is cheap, it scales, and it is the default. But similarity is not support: retrieval finds whatever resembles the question, whether or not it bears on the claim. Nothing is ever forgotten, so a superseded finding is retrieved as confidently as the correction that replaced it. Piling retrieved text into the model's context has its own cost, too: a 2025 study by Chroma across 18 models found accuracy degrading well before context limits, and fastest when the distractors were semantically similar, which is what a vault full of near-duplicate research notes provides.

2 · Auto-memory

Chat assistants like ChatGPT quietly write down "facts" about you as you talk and inject them into later conversations. Zero effort, and useful for preferences. The leak is that the curation is invisible: you do not see what was saved, what was rewritten, or what was silently dropped. Through 2025, users reported lost and garbled memories, and the vendor moved to auto-managing the store: memory that manages itself, out of sight. For a research record, a memory with no provenance and no appeal is not a record at all.

3 · Temporal knowledge graphs

Systems like Zep's Graphiti (described in arXiv 2501.13956) extract entities and relationships into a graph where every fact carries timestamps, and a superseded fact is invalidated rather than deleted, so the graph can answer "what did we believe in January?" This deserves honest credit: it is the closest existing design to an audit trail, and the invalidate-don't-delete principle is right. The costs are real, though. Building the graph takes a model call per chunk (Microsoft's early GraphRAG indexing bills ran to tens of thousands of dollars on large corpora before cheaper variants arrived), an extraction error becomes a structural error, and the change history attributes every decision to the system. No person decided anything, and the trail cannot say otherwise.

4 · LLM-maintained wikis

The newest pattern, sketched by Andrej Karpathy in 2026 and developed into a working research stack in Decoding AI's "Your second brain is a graveyard. Make it agent memory": keep immutable raw sources in one folder, let an agent write and interlink markdown wiki pages from them, and give it a schema telling it how to merge, update, and retire knowledge. A lint pass sweeps for orphan pages, stale claims, and contradictions. This is the most researcher-shaped design yet: plain files you can open, diff, and version. The graveyard diagnosis is right, too. But look at who does the tending. The merging, correcting, and retiring is handled by the agent itself. A hallucinated synthesis, written into a wiki page, is durable: every later page builds on it. RAG's errors die with the query; a wiki's errors compound. And when a claim is retired, nothing records who decided it should be, because no one did.

What forgetting research actually shows

You might expect the fix to be automatic forgetting: give the memory a decay curve and let old notes fade. That has been tried. MemoryBank (arXiv 2305.10250) applies the Ebbinghaus forgetting curve from human memory research: memories strengthen when recalled and decay when ignored. Letta's "sleep-time compute" line has agents consolidate and clean their memory during idle time, an explicit analogy to dreaming.

A 2025 evaluation of Ebbinghaus-style pruning in LLM agents reported the trade-off plainly: memory stayed small, retrieval stayed fast, and task performance dropped. Time turns out to be the wrong signal. Age does not predict whether a note still matters; a three-year-old note holding your study's inclusion criteria is load-bearing, while last week's half-thought is compost. What predicts mattering is whether the note is used, linked, and still supported by its source, and that is a judgment someone has to make.

The pruning research garden

The gardening metaphor is older than the AI wave. Mark Bernstein wrote "Hypertext Gardens" in 1998; Mike Caulfield's 2015 talk "The Garden and the Stream" set gardens (spatial, tended, revisitable) against the chronological feed; Andy Matuschak's evergreen notes are "written and organized to evolve over time"; and Joel Hooks named the failure mode when he called his site a digital garden: left untended, everything returns to entropy, and the weeds take over.

The 2026 agent-memory wave takes that tradition seriously and then automates the gardener away. The lint-and-retire loop in an LLM-maintained wiki is pruning, aimed at the right weeds. The open question it leaves is the one that matters for research: who holds the shears?

A pruning research garden is a research vault run as agent memory with the shears handed back:

1. Every claim carries its source. A note that asserts something names what it rests on. A note with no source is a seedling: fine to keep, too weak to build on.
2. Agents do the sweep. Orphan notes, stale claims, contradictions between pages, citations that may not say what the note says they say: this is the kind of tireless, mechanical scanning agents are good at. Their output is a pile of proposals.
3. A person decides. Keep, revise, or retire, with a reason. The agent surfaces that two notes contradict each other; you decide which one your project stands on.
4. Retirement is recorded, not deleted. A pruned note moves to the trail with who retired it, when, and why: the same invalidate-don't-delete principle the temporal graphs got right, with a human name attached to the decision.

This matters more in research than anywhere else because the vault feeds the record you will be asked to defend. Citation accuracy studies in biomedicine have repeatedly found trouble at the claim-source join: one PLOS ONE analysis (Mogull, 2017) estimated that about one in seven quotations of prior work misstate the source, most of those errors major, meaning the source fails to substantiate, or even contradicts, the claim it is cited for. A memory system that silently merges and rewrites claims will add errors like these to your record. And systematic-review practice already draws the line this article draws: screening tools may rank and flag, but inclusion decisions are made by named humans, recorded per record, with reasons. Your vault should meet the standard your methods section will be held to. That is also why documenting AI-assisted analysis and keeping a research changelog pair naturally with a garden like this.

Six rules to run one

You can run a pruning research garden today, in an ordinary Obsidian vault, with or without an agent. The rules are the system:

1. Plain files, one claim per note, source attached. Markdown your agent can read is markdown you can read. Keep raw sources immutable and separate from your own synthesis; nothing, human or agent, edits a source.
2. Scope the garden to a project. A whole-life vault of thousands of notes is where every design above strains; even the wiki pattern's authors say so. A project-sized garden stays readable end to end, by you and by the agent.
3. Let agents sweep, never settle. Point the agent at the weeds: orphan notes, claims with no source, contradictions, citations that may not support their sentence. Everything it produces is a proposal. Nothing it produces is a verdict.
4. Prune by use and support, not by age. The forgetting-curve result is your warning: old is not the same as dead. Retire a note because it is unsupported, superseded, or contradicted, never merely because it is old.
5. Retire with a reason, never silently delete. Move pruned notes to a dated trail with a line on why. Your future self, your co-authors, and any reviewer can then see not just what the garden claims, but what it stopped claiming, and who decided that.
6. Date your decisions. A weekly pruning pass beats a heroic quarterly cleanup. Each pass appends to the trail, and the trail, rather than a disappearing chat log, becomes the memory of the project's judgment. The four-state method in epistemic notes gives you the vocabulary for what "keep, revise, retire" means evidentially.

What this does not do

A pruning research garden does not make your notes true, and the agents sweeping it do not verify facts. A sweep can flag that a citation may not support its sentence; whether it does is a judgment a person makes by reading both. The garden's promise is narrower and more durable: what it claims is sourced, what it stopped claiming is recorded, and every consequential change has a name and a reason attached. None of that certifies a manuscript; it gives you a record you can stand behind when someone asks how you know.

Related guides

New to the underlying ideas? The AI-concepts map explains RAG, GraphRAG, and agents from scratch, and local models, agents & orchestration covers running the machinery privately. For the vault itself, start with Obsidian for researchers and the plugin shortlist, connect your references with Zotero, and grade what your notes actually establish with epistemic notes. More guides are on the Learn page.