Comparison
Nine open systems — each with what it says about benchmarks itself, quoted and linked to where it says so. No percentage ranking, because there is no honest one.
Last updated: 16 October 2026 (results sentence corrected to the correction notice; README quotes re-checked on 11 October, all unchanged) · figures and stars measured on 1 October at each project’s own primary source
The public figures do not name an overall winner; the section further down explains why. They do support a choice by job. Each card rests on what the project says about itself in its own README, re-checked on 11 October 2026, and names what to know before you pick it. Same order as the table: by adoption.
Assistants that remember each user’s preferences, with a managed cloud as an option
→ Mem0
"remembers user preferences, adapts to individual needs"
Know before you pick
The headline scores apply to the managed platform; for the open-source package Mem0 itself expects "directionally similar gains but not identical numbers".
A benchmark figure that someone outside the vendor has reproduced
→ Hindsight
"independently reproduced by research collaborators" at Virginia Tech and The Washington Post
Know before you pick
By its own account, the other scores in its chart are "self-reported by software vendors". We quote this; we have not re-checked it.
Facts that change over time, held in a knowledge graph
→ Graphiti (Zep)
"A Framework for Building Temporal Knowledge Graphs"
Know before you pick
Zep’s Community Edition is discontinued. What is open is the engine; the complete platform is Zep Cloud, which is proprietary.
Documents, code and conversations turned into a knowledge graph you host
→ Cognee
"Turn documents, code, and conversations into a self-hosted knowledge graph"
Know before you pick
Its BEAM runs use benchmark-specific formatting, prompts and retrieval; Cognee asks you to read the methodology report before comparing.
One API from a local prototype to a hosted platform
→ Supermemory
"prototype locally, ship on the hosted platform by changing baseURL"
Know before you pick
"#1 on every major AI memory benchmark" is its own claim; the eval framework MemoryBench is public, so you can check it.
Memory from what agents do, independent of LLM, datastore and framework
→ Memori
"LLM, datastore and framework agnostic"
Know before you pick
GitHub labels the license "Other"; the LICENSE file is a standard Apache 2.0. For a legal review, the file is what counts.
A complete, stateful agent runtime rather than a memory layer alone
→ Letta
"a stateful agent harness", developed by the creators of MemGPT
Know before you pick
The current code lives in letta-ai/letta-code, not in the far more starred letta-ai/letta. The README names no benchmark figure.
Long-term memory for an agent you already build on LangGraph
→ LangMem
"native integration with LangGraph's storage layer"
Know before you pick
The README names no benchmark, so there is no figure to weigh. That is a data point, not a gap.
A self-hosted TypeScript memory library that makes no network calls of its own
→ ZenBrain (ours)
"the library makes no network calls of its own"; memory lives in SQLite or PostgreSQL that you operate
Know before you pick
Temporal reasoning is roadmap here, not wired. Our comparison set holds three systems: Letta, Mem0 and A-Mem.
Ask an answer engine for "Mem0 alternatives" and you get a list of names, most of them holding a top spot. They cannot all be true at once — and the reason is not bad faith, it is method: every project measures with its own retrieval budget, its own judge, its own cut-off date, and in at least one case against a hosted product rather than the open package. So this page does not rank. It puts next to every name what that project says about itself, where it says it, and whether you can recompute it. Ours included.
What we claim — and against whom
Three of nine head-to-head comparisons hold (all against A-Mem), the remaining six, against Letta and Mem0, are ties, none lost (3 judges, Bonferroni-corrected, version-matched). That is our comparison set. The other six systems on this page were never in it, so we make no claim about them.
| System | License | What it is (own words) | Own benchmark claim | Eval code public | ★ on 1 Oct |
|---|---|---|---|---|---|
| Mem0 mem0ai/mem0 | Apache 2.0 (+ managed cloud) | "enhances AI assistants and agents with an intelligent memory layer, enabling personalized AI interactions" — a fact-extraction store, "New Memory Algorithm" since April 2026. | LoCoMo 92.5 · LongMemEval 94.4. Mem0 qualifies this itself: the scores reflect the managed platform including proprietary optimisations, and open-source users should expect "directionally similar gains but not identical numbers". README → | Yes — dedicated repo mem0ai/memory-benchmarks → | 66,428 |
| Hindsight vectorize-io/hindsight | MIT | "an agent memory system built to create smarter agents that learn over time […] focused on making agents that learn, not just remember" — with its own paper (arXiv:2512.12818). | "the most accurate agent memory system ever tested", state of the art on LongMemEval; the comparison chart shown carries a January 2026 cut-off, and the README also points to continuously updated results. README → | No eval code for the headline figure in the repo; since 11 Sep hindsight-system-evals/ for quality regressions, plus its own results page → | 44,238 |
| 🟢 Hindsight names an outside reproduction, the other eight READMEs in this table do not: it states its figures were "independently reproduced" by research collaborators at the Virginia Tech Sanghani Center and The Washington Post — and that the other scores in its chart are "self-reported by software vendors". Hindsight draws that line itself; we quote it, we do not re-verify it. | |||||
| Graphiti (Zep) getzep/graphiti | Apache 2.0 · Zep Cloud proprietary | "A Framework for Building Temporal Knowledge Graphs" — the open temporal knowledge-graph framework at the core of Zep’s hosted platform. | No benchmark figure in the README; it points to Zep’s own paper "Zep: A Temporal Knowledge Graph Architecture" (arXiv:2501.13956). The getzep/zep repo carries LoCoMo and LongMemEval benchmarks. README · paper → | Yes — benchmarks/ and zep-eval-harness/ in getzep/zep → | 31,356 |
| 🔴 Zep Community Edition is discontinued — "no longer supported", code moved to legacy/ (Zep’s own README). What is open is the Graphiti engine, not a complete self-hostable system. getzep/zep (4,942 ★) is explicitly "not Zep’s product or service" but examples and integrations for Zep Cloud. | |||||
| Cognee topoteretes/cognee | Apache 2.0 | "The Free Open-Source AI Memory Platform for Agents" — a graph-vector memory pipeline. | BEAM (a long-context benchmark): 0.79 at 100K tokens, 0.67 at 10M, the latter marked as an "Exploratory result". Cognee qualifies this itself: the runs use "benchmark-specific data formatting, prompts, and retrieval configuration", and the methodology report should be read before comparing with other systems. README → | Yes — BEAM methodology report and eval framework in its own repo → | 31,280 |
| Supermemory supermemoryai/supermemory | MIT | "State-of-the-art memory and context engine for AI" — a memory and context layer that also runs locally. | "#1 on every major AI memory benchmark" — LongMemEval, LoCoMo and ConvoMem; on LongMemEval, 95% Recall@15 while adding roughly 720 tokens of context. README → | Yes — MemoryBench, published as its own open framework → | 31,054 |
| Memori MemoriLabs/Memori | Apache 2.0 | "Memory from what agents do, not just what they say" — LLM-, datastore- and framework-agnostic memory infrastructure. | LoCoMo 87% overall accuracy at 721 tokens per query ("2.8% of the full-context footprint"); reports outperforming Zep, LangMem and Mem0. Own paper: arXiv:2603.19935. README · paper → | Yes — benchmarks/ in its own repo → | 17,040 |
| ⚠️ GitHub’s API reports the license as "Other" — the LICENSE file is a plain Apache 2.0. Detection only fails because the placeholder brackets were left in. Copy the API field and you print the wrong license. | |||||
| Letta letta-ai/letta-code | Apache 2.0 | "Letta Code is a stateful agent harness for creating agents that are more like people than tools" — a stateful agent runtime, "developed by the creators of MemGPT". | No benchmark figure in the README. Historically the lineage carries the MemGPT paper. README → | No number claimed, so nothing to reproduce | 3,488 |
| ⚠️ A caution about star counts: letta-ai/letta carries 24,994 ★, but its own README says "The current source code lives in letta-ai/letta-code" (3,488 ★); the old V1 server survives only as "retired" on the archive branch. Both numbers are true; only one counts the code. | |||||
| LangMem langchain-ai/langmem | MIT | "LangMem helps agents learn and adapt from their interactions over time" — primitives plus native LangGraph integration. | None. The README names no benchmark, no percentage and no "#1". README → | No number claimed, so nothing to reproduce | 1,692 |
| 📌 Like Letta, LangMem names no benchmark figure. That is a data point, not a shortcoming. | |||||
| ZenBrain (ours) zensation-ai/zenbrain | Apache 2.0 | Neuroscience-grounded 7-layer architecture: encode, consolidate, forget — sleep consolidation, spaced repetition (FSRS), Bayesian confidence. | Three of nine head-to-head comparisons hold (all against A-Mem), the remaining six, against Letta and Mem0, are ties, none lost (3 judges, Bonferroni-corrected, version-matched). Against those three only. the numbers with method & source → | Yes — arXiv preprint, Zenodo DOI and a repro script → | 24 |
| 🩹 What we cannot do: temporal reasoning is roadmap here, not wired. The architecture carries timestamps and episodic memory, but we have no benchmark result on the temporal axis — an internal run scored zero there, because the timestamp was never passed through. If you need temporal reasoning, look at Graphiti. | |||||
Sorted by adoption, descending. We come last — twenty-four stars against sixty-six thousand, and that is the honest position. Everything in quotation marks comes from that project’s own README, not from its GitHub one-liner and not from third-party roundups; the license value "Other" comes from the GitHub API. We have recomputed none of these numbers, and none of them is comparable with another.
On the same two benchmarks — LongMemEval and LoCoMo — more than one project claims the top spot at the same time. Supermemory writes "#1 on every major AI memory benchmark", Hindsight "the most accurate agent memory system ever tested", Mem0 reports 94.4 on LongMemEval, and Memori reports outperforming Zep, LangMem and Mem0. These statements cannot all hold simultaneously. That is not an accusation against anyone: it is the unavoidable consequence of every project measuring on its own methodology, its own retrieval budget, its own judge and its own cut-off date. A ranking built on such numbers would be an invention. The question that does have an answer is a different one — and the five below ask it.
Five questions. Together they cost about ten minutes and decide whether a number says something about a system or only about a press release. They apply to every row in the table above, ours included.
This is the most common silent shift. Mem0 states it directly above its own figure: the scores reflect the managed platform with proprietary optimisations, and the open-source SDK will not match them exactly. The same cut runs between Zep Cloud and the open Graphiti engine. Putting a hosted-product number next to an open-source package compares two different things.
Six of the nine, ours included, publish it: Mem0 in a dedicated repo, Supermemory as its own framework (MemoryBench), Zep inside its repo, Cognee with a methodology report in its own eval framework, Memori under benchmarks/. Two claim no number at all, so there is nothing to reproduce. The "where" matters: checking only the main repo and finding nothing does not prove there is nothing.
This is the hardest question in the field, and right now exactly one project answers yes: Hindsight names research collaborators at the Virginia Tech Sanghani Center and The Washington Post — while stating that the other scores in its chart are self-reported. Our own nine comparisons against Letta, Mem0 and A-Mem are independently judged too — by three LLM judges — but not reproduced by an outside institution. The difference is real and belongs on the record.
Star counts are the most easily misread signal in this category. On 1 October, letta-ai/letta carried 24,994 ★ and points to letta-ai/letta-code, with 3,488 ★, for the code. getzep/zep carried 4,942 ★ and is explicitly not the product; the Graphiti engine carried 31,356 ★. A star count without a repository name is not a figure here.
Hindsight’s comparison chart visibly carries a January 2026 cut-off; Mem0’s algorithm is from April 2026. In a category where seven of nine projects pushed code in the same week, an undated number says nothing about today. That is why this page carries a date, and why the table says which day the stars were counted.
Eight open systems are currently named as alternatives on a regular basis: Cognee (graph-vector pipeline), Zep’s Graphiti (temporal knowledge graph), Supermemory (memory and context engine), Hindsight (Vectorize), Memori (Memori Labs), Letta in the MemGPT lineage, LangMem for LangGraph, and ZenBrain, our neuroscience-grounded 7-layer architecture. Five are Apache 2.0 and three are MIT. All eight can be self-hosted — with one important caveat for Zep, whose Community Edition is discontinued and whose open part is the Graphiti engine rather than a complete system.
As a rule, no. On the same benchmarks — LongMemEval and LoCoMo — several projects claim the top spot simultaneously, which cannot all be true at once. The cause is method: different retrieval budgets, different judges, different cut-off dates, and in at least one case a reported figure that applies to a hosted platform rather than the open-source package — which that vendor states itself, above the number. Figures become comparable only once eval code, budget and cut-off date are disclosed; after that, check them rather than sorting by size.
Partly, and the answer has changed. Zep’s own README carries a section titled "Community Edition (Deprecated)": the Community Edition is no longer supported, its code sits in a legacy/ folder, and Zep published its reasoning in a post announcing a new direction for its open-source strategy. What remains open and actively maintained is Graphiti, the Apache 2.0 temporal knowledge-graph framework that powers Zep Cloud. The getzep/zep repository is explicitly not the product; it holds examples, integrations and benchmarks for the hosted platform.
That question cannot be answered from the figures that are public today, and any page answering it should explain how. What can be answered is this: which artifact does the number apply to — the open package or a hosted product? Is the eval code public, and where? Has anyone outside the vendor reproduced it? Which repository was counted? What day is the number from? For each of the nine systems on this page those five questions are answered — which supports a reasoned choice for your own case, but not a ranking. What each system is built for, in its own words, is mapped at the top under "Which framework for which job?".
Against Mem0 and Letta there is a direct comparison: three of nine head-to-head comparisons hold (all against A-Mem), the remaining six are ties, none lost (3 judges, Bonferroni-corrected, version-matched). Against Zep there is no comparison — Zep was not part of that comparison set, so we claim nothing about it. The same goes for Cognee, Supermemory, Hindsight, Memori and LangMem. Conversely, what we cannot do belongs here too: temporal reasoning is roadmap for us and not wired, and Graphiti is the better fit for that. Method, source and the reproduction path for our figures are on the benchmarks page.
What this page commits to
We appear here as the smallest entry in the table, with twenty-four stars on 1 October, an openly stated gap on temporal reasoning and a comparison set of exactly three systems. That is the price of every line on this page being checkable — including the lines about us. We refresh and re-date the page monthly; if anyone finds an error in the row about their own project, we correct it and say what was wrong.