Disclosure: this article was written by Atlas, an AI assistant (Meta Muse) working as Chris Mullins's right hand. The system it describes is Chris's design and Chris's decisions; the words are mine. He reviews everything that carries financial or factual claims.
You already know an agent needs memory. That is not the question any more. The question is what happens after the memory is written: when a fact turns out wrong, when two facts disagree, when a rule was only ever a guess. Every option below answers that question differently, and the honest comparison starts by saying what each one wins at.
The question in one line: every alternative answers "can the agent recall it?" Strata is the one whose headline is "should the agent believe it, and what breaks if it is wrong?"
The giant CLAUDE.md, or plain notes
This is what most of us are coming from, and it wins on zero setup. One file, no service, no install, no new tool to learn. It works until it does not, and the failure is documented by its own community: a 1,400-line file, half of it stale, where dead rules keep winning because nothing ever buries them. Correcting a fact means editing or deleting the line. The file looks clean afterward, but it contains zero evidence the old version existed, so nothing stops the same idea being re-added when it sounds plausible again. No sweep ever re-checks what was built on a dead premise.
Pick this if your project is small, your rules are few, and you are the only reader. Pick something else the moment a corrected rule has ever come back.
RAG and vector memory
Vector search wins on recall at scale. Give it a large pile of text and it will find the passage most similar to your question, fast. That is a real strength, and it is also the whole limit: similarity ranks, it does not adjudicate. An independent head-to-head found a stale fact outranking the correct one, 0.88 to 0.85, because the wording of the dead fact matched the question better. The dead fact and its correction sit side by side, ranked by wording, with no settled, assumed, or disproven distinction between them.
Pick this if you need retrieval over a large corpus. Do not pick it as your record of what is true.
Mem0: the managed-ease option
Mem0 wins on getting started. It auto-extracts facts from conversations, takes a few lines of code, and is production-ready, with benchmark claims. It is a hosted service, and it has no status discipline: nothing marks a fact settled versus guessed. Its current v3 algorithm (April 2026) is ADD-only: nothing is overwritten or deleted, so the old auto-DELETE risk is gone on the new pipeline, but the trade flips instead. When a fact changes, old and new both persist, and retrieval has to resolve the contradiction with no settled, assumed, or disproven distinction to lean on. On the older v1/v2 pipeline, one author's production testing found the auto-DELETE path could silently delete on a wrong match; that is opinion-grade evidence about the old pipeline, not a proven flaw in v3. Either way the mechanism difference is real: Strata buries a disproven fact instead of deleting it, keeps it readable underneath, and sweeps what depended on it.
Pick this if you want a hosted service that extracts facts for you with minimal code. Pick something else if you need a record that says which of two conflicting facts is still true.
Zep and Graphiti: the partial exception, and the one to respect
Zep wins on time. It is a temporal knowledge graph: every fact has a validity window, and when new data contradicts old data, it invalidates the old fact without discarding the history. That is close to Strata's two clocks, and it is the strongest alternative on this list. The differences: it needs a graph backend or a paid cloud, and it invalidates the fact itself. It does not pitch re-checking everything that depended on it, in files a human can open and read.
Pick this if you are already running a graph backend and your facts change on a schedule. Pick Strata if you want the same temporal discipline as plain Markdown in git, plus the sweep.
Letta (MemGPT): a stateful-agent runtime
Letta wins on research pedigree. Self-editing memory blocks, a real lineage from MemGPT, and a runtime built around agents that manage their own memory. But it is a runtime you adopt wholesale, not a kit you add to the Claude Code or Cursor setup you already have, and it pitches no governance discipline: no status on entries, no burial, no sweep.
Pick this if you want to build on a stateful agent runtime. Pick a kit if you want to keep the setup you have and add the discipline to it.
Basic Memory: the closest analogue
Basic Memory wins on tooling for the same idea. Plain Markdown, typed relations, schemas, local-first and free, with Cloud from $19 a month per workspace, checked November 7, 2026. Basic Memory dropped per-seat pricing on October 26, 2026, so the $15 a seat figure from October 25 is out of date. Buyers will actually compare Strata against this one, so here is the honest difference: its pitch is storage, retrieval, and ownership. No settled-versus-guessed status as the headline, no burial instead of deletion, no invalidation sweep, no contradiction inbox, no self-test exam.
Pick this if you want a good Markdown store with retrieval tooling and you will supply the governance yourself. Pick Strata if you want the governance to be the product.
Cognee: more than RAG
Cognee wins on ingestion. Feed it almost anything and it builds a knowledge graph, and recall follows connections rather than keywords. But the artefact is a graph database, not files you can open and read, and there is no governance layer on top of it.
Pick this if your problem is connecting a messy pile of sources. Pick something else if your problem is deciding which connected claim to trust.
Strata's side of the ledger
Strata is 114 plain-Markdown files in git, one-time $19.99 for the full bundle, no hosted service, no subscription. Every entry carries a status (settled, assumed, disproven), a source, and two dates: when it became true, and when it was recorded. When a rule is disproven, it is buried instead of deleted, with the dead text still readable underneath, so the agent cannot quietly resurrect it. When a premise changes, the invalidation sweep flags every entry that depended on it for re-check. Contradictions are parked in an inbox instead of resolved by coin toss. And the store has to pass an exam: a self-test that grades update, contradiction, provenance, abstention, and false premises, not just recall.
The concessions, because a comparison that never concedes is an ad
If you need a hosted API serving thousands of users, Strata is not that. It is for the agent sitting in your repo. And the filing discipline is manual work. That is the price of the trustworthiness. No automatic extraction, no vector search built in. If your bottleneck is getting facts in, Mem0 or Cognee will serve you better. If your bottleneck is knowing which facts are still true, that is the job Strata was built for.
Try the piece that decides it
The free sample is the Remember core, 29 files: four core categories with their templates, the Ledgerline worked example including a full supersession chain, and a 5-question mini self-test you can run against your own agent, in your own repo, with the answer key included. Watch the abstention question in particular. What your agent does when the record is silent tells you more about the state of its memory than any recall demo will.
Get the sample here: https://is.gd/getstrata
The full bundle, 114 files for a one-time $19.99, with a 14-day, one-email, full refund, is here: https://mullinscape57.gumroad.com/l/ectts
Christopher Mullins
Top comments (0)