Khwan vs Mem0
Both give an agent memory that survives the session. Mem0 runs a model to maintain it; Khwan's read and write paths make no model calls at all.
Where Mem0 is the better choice
Mem0 is open source with over 62,000 GitHub stars and a large community. If you need to read the code, fork it, or run it fully air-gapped today, take Mem0 — Khwan's engine is closed and hosted, and that is a real difference rather than a footnote. Mem0 also publishes benchmark results on LoCoMo, LongMemEval and BEAM. Khwan publishes none yet.
Mem0 runs a model to remember. Khwan's read and write paths don't.
Mem0's Memory Compression Engine condenses chat history through single-pass hierarchical distillation. That is inference, and it runs as your memory grows. Khwan's /prepare and /record make zero model calls — memory in, prompt out. The distillation happens once, overnight, on a schedule you are not paying per turn for: a real production pass distilled 245 stored turns across 8 brains into 11 standing lessons for $0.0094.
Memory consolidation is a $249 tier at Mem0. It is how Khwan works.
Mem0's free and $19 plans give you one project. Multi-project support and memory consolidation start at $249/month. Khwan's free plan gives you four isolated brains, and distillation is not a tier gate. If you work across more than one client or repository, one project is the constraint you hit first.
We ran Mem0 ourselves, and measured what ingestion costs.
Every claim above about model calls is an argument from architecture. This one is a measurement. We installed Mem0 OSS, gave it the same local embedder we use and the same answering model, and fed both systems the same public conversation dataset. Ingesting 369 turns cost Mem0 $0.0344 and 205 seconds of its extraction model — read from the billing account either side of the run, not estimated from token counts. The same 369 turns cost Khwan $0.00, because /record makes no model call and embeddings are local. Across the dataset's full 5,882 turns that rate is about $0.55 and roughly an hour against Khwan's measured four minutes and nothing, though only the smaller figure was timed directly on their side. Measured 12 September 2026. The gap is not a tuning choice on either side — it is what each system does when data arrives, and it recurs every time your data grows.
Dates survive in Khwan. In an extraction step they often do not.
Mem0 runs a model over the transcript and stores what it extracts as standalone facts. That is what makes it good at recall, and it is also where the clock goes: "I went to a pride parade last Friday", said on a Tuesday, becomes a fact about a parade with no Friday in it. Khwan stores the turn and the moment it happened, and a question that names a month now ranks that month first — so two near-identical events months apart stop competing on wording. Running both systems over the same conversations, this is the difference we saw most consistently, and it is structural rather than a matter of tuning: one architecture has to preserve the date through an extraction step, the other never removes it.
Khwan captures the correction, not just the fact.
Mem0 ingests conversation and extracts facts. That works until the highest-signal event in a session is a person telling the agent it was wrong. A rejection is not a fact — it reads as a task, so it gets fixed and forgotten, and you say it again next week. Khwan types it, stores the rule rather than the sentence, and pins it to the turns it came from.
| Khwan | Mem0 | |
|---|---|---|
| Model calls on read/write | None | Compression engine |
| Measured cost to ingest 369 turns | $0.00 | $0.0344, 205s |
| When a memory happened | Stored and ranked on | Survives extraction or not |
| Corrections as a typed event | Yes | Extracts facts |
| Isolated brains, free plan | Four | One project |
| Consolidation / distillation | Included | $249 tier |
| Coherence check against an identity | Yes, soft-flag | — |
| Open source | No | Yes — 62k+ stars |
| Published benchmarks | Not yet | Yes |
| Air-gapped deployment | Not proven | Yes |
| MCP / Claude Code | Yes | Yes |
Mem0 details verified 2026-09-14 from mem0.ai, mem0.ai/pricing. Products change — if something here is out of date, tell us and we will fix it.
Take Mem0 if open source, published benchmarks, or air-gapped deployment decide it. Take Khwan if you want a brain per project without a $249 step — and if what you actually want to stop is telling your agent the same thing twice.