Mr. MeeseeksThe numbers, and their limits
Research
What we have measured so far, how we measured it, and where the claims stop.
Memory
Better than Mem0 at every level of model intelligence. On a 310-question LoCoMo memory benchmark (top 10 retrieved), Mr. Meeseeks beat Mem0 with every answer model we tried.
| Answer model | Mr. Meeseeks | Mem0 |
|---|---|---|
| GPT-4o | 91.3% | 77.4% |
| Sol | 92.3% | 80.7% |
| Sonnet 5.5 | 93.9% | 84.5% |
Your memory is stored in local files on your machine.
Honest limits: this is a 310-question slice, not Mem0's full 1,540-question set, and our memory lookup is slower than Mem0's (about 5.0s vs 0.8s).
Context management and compaction
Better caching and less context means far cheaper usage. Mr. Meeseeks keeps the AI's working memory small and stable, so roughly 80% of what you would otherwise pay for model input is avoided.
Instead of resending an ever-growing chat history on every turn, older conversation is folded into layered summaries written by a cheap model (Haiku). The agent can “zoom” back to the exact original messages whenever it needs detail, and always works from a compact view (about 90–130KB) that doesn't keep growing.
- Far cheaper turns. Because the context stays stable, providers can reuse cached input instead of charging full price. Across about 1,300 real turns, most input came from cache: Opus 97%, Sol 94%, Astra 94%, Haiku 93%, Sonnet 86%, Terra 82%.
- Cost doesn't balloon. The working memory has a ceiling, so a months-long project costs about the same per turn as day one.
- Nothing is lost. Summaries point back to the originals, so the agent recovers exact details instead of guessing.
Subscription and API cost management
Across all logged usage, caching cut the API-equivalent cost from about $5,200 to about $960: roughly $4,239 saved, an 81.5% reduction.
| Provider | Usage | Cache rate | With cache | Without cache | Savings |
|---|---|---|---|---|---|
| Claude turns | 525 turns | 94.3% | $239.46 | $1,246.85 | $1,007.39 |
| Claude summaries | 26,758 calls | 65.6% | $51.81 | $88.72 | $36.91 |
| Claude total | $291.26 | $1,335.57 | $1,044.31 (78.2%) | ||
| GPT/Codex total | 785 turns | 93.0% | $668.93 | $3,863.99 | $3,195.07 (82.7%) |
| Combined | 1,310 turns + summaries | $960.19 | $5,199.56 | $4,239.38 (81.5%) |
Honest limits: these are estimates from our own usage logs at published list prices, not invoices. Most of our usage runs on subscriptions, so this is not money actually billed.
Infinite agents
- A worktree per child. Each child gets its own git worktree. The parent's slot frees up when it delegates, and the parent merges the child's branch afterward.
- Controllable jobs. Every job has a #N you can zoom into, tell, or stop, so the work is visible, steerable and durable.
- Mixed vendors. Use Claude and GPT models together; Claude and Codex subagents stay inside their own vendor.
- Memory for children. Children can zoom back to the original messages instead of working from summaries.
- Spawn receipts. “Started” is only claimed when a spawn receipt exists.
Owned data
Your memory lives in local files that you own.
No vendor lock-in
Use Claude and GPT models side by side, and switch between them without losing your memory.