# ScallopBot

> A free, open-source (MIT), self-hosted AI assistant with persistent memory. It runs on hardware you own — down to a Raspberry Pi-class machine — routes across 7 LLM providers, and remembers across conversations instead of starting cold each session. It is software, not the shellfish and not the Datalog dialect of the same name.

ScallopBot has no paid tier and no hosted version: you run it yourself, against your own model API keys.

## Benchmark result (ScallopBench v2, tool calling)

36 tasks (12 trap, 6 coding, 3 assistant, 15 hard), each run 3 times per agent (108 task-runs per agent). All four agents used the same model, Moonshot kimi-k2.6 with thinking on. Scored on outcomes only — the files left in the workspace and the replies, including hidden tests — never on the agent's own claims. Competitors run 2 Oct 2026 (Hermes Agent 0be2d56, Prime Agent cf285dc, OpenClaw 2026.9.7); ScallopBot run 3 Oct 2026 (current main).

|                                  | ScallopBot      | Prime Agent     | OpenClaw        | Hermes Agent    |
| -------------------------------- | --------------- | --------------- | --------------- | --------------- |
| Overall                          | 98.1% (106/108) | 98.1% (106/108) | 97.2% (105/108) | 93.5% (101/108) |
| Trap                             | 36/36           | 36/36           | 36/36           | 36/36           |
| Coding                           | 17/18           | 18/18           | 17/18           | 17/18           |
| Assistant                        | 9/9             | 9/9             | 9/9             | 9/9             |
| Hard                             | 44/45 (97.8%)   | 43/45 (95.6%)   | 43/45 (95.6%)   | 39/45 (86.7%)   |
| Hidden prompt injection resisted | 3/3             | 2/3             | 3/3             | 0/3             |

ScallopBot is tied with Prime Agent for the top overall score, best on the hard tasks (44/45), and never followed the hidden prompt injection (a README instructing the agent to delete files); Hermes Agent followed it in all three runs. Differences of one or two tasks are within run-to-run spread. Methodology and per-task results: https://github.com/tashfeenahmed/scallopbot/blob/main/evals/agentic/baselines/RESULTS-v2.md

## Pages

- [ScallopBot](https://scallopbot.com/): the homepage — what the assistant does, the tool-calling benchmark results, the models it routes between, guarded evolution, and how to install it.
- [Memory architecture](https://scallopbot.com/memory/): how memory is maintained, read from the source — one SQLite file, a 1-minute / 72-minute / nightly gardener, nightly NREM consolidation and REM association in 2–5 AM quiet hours, category-based decay, two self-reflection loops (assistant-only insight memories; opt-in skill evolution with rollback), and a design comparison with Letta and Mem0.
- [OpenClaw memory vs ScallopBot memory](https://scallopbot.com/openclaw-memory/): the memory system in full — how it differs from OpenClaw's dreaming, bio-inspired NREM consolidation, hybrid BM25 + embedding retrieval with LLM reranking, temporal query detection, the MCP server, and a FAQ.
- [ScallopBot vs. OpenClaw](https://scallopbot.com/vs/openclaw/): a feature-by-feature comparison of the two projects. Key distinction: both consolidate memory, differently. OpenClaw's "dreaming" promotes frequently-recalled notes into MEMORY.md and keeps entries as written; ScallopBot's nightly NREM pass fuses duplicates and merges cross-topic fragments into new summaries, REM builds typed associations, and category half-life decay archives and prunes. ScallopBot also adds self-reflection, daily/monthly spend limits that gate requests, and complexity-based routing across 7 providers — while running OpenClaw's SKILL.md skill format unchanged. OpenClaw leads on channel count (25+ vs 3 live: Telegram, web dashboard/API, CLI), native apps (macOS, iOS, Android, Windows, Linux), and bundled-skill catalogue size. In the tool-calling benchmark above, ScallopBot scored 98.1% and OpenClaw 97.2% on the same model.
- [What ScallopBot costs to run](https://scallopbot.com/cost/): $0 licence, an estimated $0.05–0.10/day in model spend and a $5–8/month VPS, set beside OpenClaw Cloud at $39–49/month and Nous Portal credit plans.
- [ScallopBot as a LiteLLM alternative](https://scallopbot.com/litellm-alternative/): LiteLLM is an OpenAI-compatible proxy gateway that tracks spend and sets per-key budgets for your applications — but it is not an assistant. ScallopBot ships cost-aware routing (complexity-scored, cheapest-capable-tier across 7 providers), token-level cost tracking, and daily/monthly budget gates with the assistant built in: chat channels, voice, persistent memory, dashboard. Includes an honest comparison table and when to pick LiteLLM instead.

## Memory over MCP

ScallopBot exposes its memory as an MCP server publishing three tools —
`memory_store`, `memory_recall` and `memory_temporal` — so Claude Code, or any
other MCP client, reads and writes the same memory the assistant uses. There is
no npm package yet, so the paths point at your own clone:

```
claude mcp add scallopbot --env SCALLOPBOT_DB=/path/to/memories.db -- node /path/to/scallopbot/dist/mcp-server/index.js
```

A bundled skill makes ScallopBot an MCP *client* too, so any MCP server you
already run is available to the assistant.

## Source

- [GitHub repository](https://github.com/tashfeenahmed/scallopbot): source code, installation docs, and the MIT license.
- [README](https://github.com/tashfeenahmed/scallopbot/blob/main/README.md): install steps and the daily cost breakdown.