cogito-memResults
Is Cogi any good?
We tested it the way memory services test themselves, on a laptop, with the model that ships. Short answer: as good as rereading everything, for $0, in milliseconds.
of LoCoMo questions answered right from 1,000 tokens
of what rereading the whole conversation scores
to search 10 million tokens, laptop CPU only
AI calls to save a message, so $0 per message
Answers, next to the others
On LoCoMo (10 long conversations, 1,540 questions), an AI model answers each question from what the memory hands it, and a second AI model grades the answer. This is the test Mem0's paper used for Mem0, Zep and others.
How close each memory gets to rereading everything
Score as a share of the whole-conversation score measured the same way · higher is better
Others' scores are from the Mem0 paper. Zep disputes Mem0's run of Zep; in its own rerun Zep reports 75.1, which is 103% by this measure.
The numbers in a table
| Memory | Score | Whole conversation | Share | Source |
|---|---|---|---|---|
| Cogi, 1,000 tokens | 74.7 | 76.5 | 98% | measured 19 Oct 2026 |
| Mem0 with graph (Mem0g) | 68.4 | 72.9 | 94% | Mem0 paper |
| Mem0 | 66.9 | 72.9 | 92% | Mem0 paper |
| Zep, run by Mem0 | 66.0 | 72.9 | 91% | Mem0 paper |
| Zep, own rerun | 75.1 | 72.9 | 103% | Zep blog |
| Plain search (RAG), best setting | 61.0 | 72.9 | 84% | Mem0 paper |
| ChatGPT memory | 52.9 | 72.9 | 73% | Mem0 paper |
Why shares? Our whole-conversation run scored 76.5 where the paper's scored 72.9, so our setup may grade a little more generously. Dividing each score by the whole-conversation score measured alongside it removes that difference.
A bigger basket, better answers
Answers right, by how much Cogi hands back
LoCoMo, first 5 conversations (762 questions), graded as above
Claude gets up to 300 tokens with each prompt by itself, and 1,000 when it searches.
Cogi shines on single facts (84.5% right at 1,000 tokens) and on dates, since every message carries its own. It is weaker on questions that need many facts at once (61.7%), where a fixed basket runs out first.
Finding the right messages
Every message an answer needs, handed back within 1,000 tokens
LoCoMo, all 10 conversations, 1,536 questions
Searching by meaning and by words together beats words alone by 12.7 points (95% interval 11.0 to 14.5).
Fast
95% of searches in 10 million tokens, CPU only (target: 100 ms)
the same with the laptop's GPU
to save a message; findable by meaning a moment later
| Published search times | 95% of searches within |
|---|---|
| Cogi, 10 million tokens, laptop CPU | 26 ms |
| Zep, its own 2026 page | 155 ms |
| Mem0, open source (Mem0 paper) | 200 ms |
| Mem0 with graph (Mem0 paper) | 657 ms |
| Mem0 platform, April 2026 run | 6.0 s |
These compare designs, not hardware: hosted memories include the network and their own AI calls.
Free
Saving a message costs one local embedding: no AI calls, no API key, $0. Services that extract facts call an AI model for every message they store, and you pay for each call.
Should you pick Cogi?
Yes, if you
- work in Claude Code or Codex and lose context in long sessions or across days
- want your conversations to stay on your computer
- want no account, no API key and no bill
- want exact words and dates, not an AI's summary
Something else, if you need
- memory in cloud agents or claude.ai today
- one memory shared between machines or a team
- answers that join many facts across many sessions
- a whole agent framework (Letta), not just memory
How we measured
- Where: MacBook Pro, Apple M5 Max, 19 Oct 2026, with EmbeddingGemma 2 (Q8_0), the model that ships.
- Answers: gpt-4o-mini answers and gpt-4o-mini grades, with the Mem0 paper's prompts word for word. One run per row (about ±2.2 points). The whole conversation was run on the first 5 conversations only.
- Finding messages: a question counts when every turn LoCoMo labels as evidence is in what Cogi hands back.
- Speed: 1,000 searches one at a time, over 10 million tokens of synthetic text.
- Newer scores of 90% and more that some services publish use stronger answering models and more lenient graders; they are not comparable with these.
Everything is reproducible with cogito-mem eval and cogito-mem bench; the
full report has every number, chart and caveat.