
Mammoth
- 2 repo stars
- Updated August 3, 2026
- all-source-os/all-source
Durable agent memory. Your coding agent remembers decisions, task history, and code context across sessions and recalls them automatically. caveman make few token, mammoth never forget token.
About
mammoth is a Claude Code skill in the Automation & Workflows category. Durable agent memory. Your coding agent remembers decisions, task history, and code context across sessions and recalls them automatically. caveman make few token, mammoth never forget token.
- mammoth
- Automation & Workflows
- AI-coding skill
Mammoth by the numbers
- Data as of Aug 4, 2026 (Skillselion catalog sync)
/plugin marketplace add all-source-os/all-source/plugin install mammoth@mammothAdd your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| repo stars | ★ 2 |
|---|---|
| Last updated | August 3, 2026 |
| Repository | all-source-os/all-source ↗ |
What it does
Durable agent memory. Your coding agent remembers decisions, task history, and code context across sessions and recalls them automatically. caveman make few token, mammoth never forget token.
README.md
mammoth 🦣
Durable memory for your coding agent. Decisions, task history, and code context from past sessions get remembered and recalled automatically — so you stop re-explaining your project every single time. Works locally with zero signup (your memory, your disk, survives reboots), and syncs across machines when you want it to. Any MCP agent — Claude Code, Cursor, Cline, Windsurf, Codex, Gemini — speaks it natively.
caveman make few token. mammoth never forget token.
mammoth is the distribution front-door for two AllSource engines:
allsource-prime (semantic recall —
graph + vector + temporal over a durable event store) and
chronis (cn, event-sourced task/work
memory). caveman compresses what you say; mammoth remembers what you said.
They compose.
The magic moment
Sessions later, you ask "why did we pick X over Y?" — and the agent answers from memory, without you pasting any history. It recalled a decision from a session it was never in. That's the whole product.
What you get
- A
mammoth-memoryskill that runs the recall/record loop automatically. - A
voice-identityskill +/voicecommand — build a durable, queryable voice file (how you think, write, and decide) on the same Prime store, and have any MCP agent write in your voice by recalling the relevant slice. Replaces the viral "100-question interview → static 4k-token markdown blob" pattern with something live, portable, and team-shareable. Seedocs/proposals/PRIME_VOICE_FILE.mdand the proof intooling/voice-demo/. /remember,/recall,/memory-statusslash commands (explicit escape hatches).- The Prime MCP server wired via
.mcp.json— 13prime_*tools (prime_recall,prime_context,prime_add_node,prime_embed,prime_stats, …) plus theprime://auto-contextresource: a compressed knowledge index auto-injected into every conversation.
Prerequisite (all agents)
cargo install allsource-prime # needs >= 0.21.6 (in-process fastembed, no external service)
The binary must be on your PATH.
Install
Claude Code (richest — plugin)
/plugin marketplace add all-source-os/all-source
/plugin install mammoth
Restart and approve the project prime MCP server when prompted. The
mammoth-memory skill then activates automatically.
One-line installer (any supported agent)
curl -fsSL https://raw.githubusercontent.com/all-source-os/all-source/main/plugin/mammoth/install.sh | bash
Auto-detects your agents, installs the binary if needed, and writes the MCP
config. --print to preview, --help for flags. Windows: install.ps1.
Install matrix
mammoth's core capability is one MCP server — any MCP agent gets all 13 tools by adding one stanza. See INSTALL.md for per-agent placement.
| Agent | Path |
|---|---|
| Claude Code | plugin (above) or project .mcp.json — also gets the skill + slash commands |
| Cursor | .cursor/mcp.json |
| Cline | Cline MCP settings |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
| Codex | Codex MCP config |
| Gemini CLI | ~/.gemini/settings.json |
The shared MCP stanza:
{
"mcpServers": {
"prime": {
"command": "allsource-prime",
"args": ["--data-dir", "~/.prime/memory", "--auto-inject", "--auto-inject-max-tokens", "1000"]
}
}
}
Does it actually work? (benchmarks)
Reproducible harness in tooling/mammoth-bench/
— run it yourself. Recall is scored against a keyword search+grep baseline over
the same store, with queries deliberately worded unlike the stored text.
Full run, 60 memories / 60 queries, k=5 (bench2.py):
| metric | result | kind |
|---|---|---|
| Recall hit@5 | 0.90 (baseline 0.83, Δ +0.07) | measured |
| Recall hit@3 / MRR | 0.87 / 0.783 | measured |
| Cross-session continuity win-rate | 0.07 | proxy |
| Median tokens saved / recall | 19 (986 total over 54 hits) | estimate |
| Recall latency p50 / p95 | 3.0ms / 3.6ms | measured |
| Durability (write→restart→read) | PASS | measured |
Honest note: on a smaller 18-memory set the edge was +0.17; it narrows to +0.07
as the corpus densifies and the keyword baseline gets more overlap to exploit.
Memory still wins every metric. The improvement levers (hybrid keyword+vector,
graph depth, a larger embedder) are unused here — headroom, not yet spent. We
publish the softening rather than hide it.
How it stores
Local-only by default: allsource-prime --data-dir <project>/.prime runs a fully
durable store — WAL (CRC32 + fsync) + Snappy Parquet snapshots, the same
AllSource Core engine. Your memory survives restarts and never leaves your
machine. No account required.
Cross-machine / team memory is the upgrade: add --sync-to <url> --api-key <key>
to ship prime.* events to a hosted AllSource Core.
Free tier remembers across sessions on this machine, no account. Cross-machine and team memory need a free AllSource account.
Provenance
mammoth records its own analysis — decisions, rationale, gotchas, outcomes, metadata, file paths. It does not write raw source code, secrets, or full file contents into memory unless you explicitly allow it.
License
MIT (plugin layer — plugin/mammoth/ + .claude-plugin/). The engines:
allsource-prime and allsource-core are Apache-2.0; chronis is MIT. See
LICENSE.