all.sourceAllSource

Your Coding Agent Has Goldfish Memory. Give It a Mammoth.

Every new session, your coding agent forgets. The architecture decision you made last Tuesday, the gotcha that cost you an afternoon, the reason you picked library X over Y — gone. So you paste it back in. Again. And again.

mammoth fixes that. It gives your agent durable memory: decisions, task history, and code context from past sessions get remembered and recalled automatically — so you never re-explain your project twice.

mammoth is the front door for two AllSource engines: allsource-prime (semantic recall — graph + vector + temporal over a durable event store) and chronis (cn, event-sourced task memory). It pairs naturally with caveman: caveman compresses what you say; mammoth remembers what you said.

caveman make few token. mammoth never forget token.

The magic moment

Here's the whole product in one interaction. Three sessions after you decided something, you ask:

"Why did we go local-only instead of requiring an account?"

And the agent answers — correctly, citing the decision — without you pasting any history. It recalled something from a session it was never part of. The first time you see it, you ask "wait, how did it know that?" That reaction is the point.

How it works

mammoth runs the AllSource Prime MCP server locally. Your agent gets 13 prime_* tools and a recall/record loop:

  • Record — when a decision, gotcha, or outcome happens, the agent writes it as a node and embeds it (prime_add_node + prime_embed). Your own analysis, never your raw source code.
  • Recall — before answering a question that depends on prior knowledge, the agent calls prime_recall / prime_context: vector similarity fused with graph neighbors and temporal recency.
  • Auto-orient — with --auto-inject, a compressed knowledge index is injected into every conversation as the prime://auto-context resource. The agent walks in already knowing the shape of what it knows.

Embeddings are computed in-process via fastembed (AllMiniLML6V2, 384-dim) — no external embedding service, no API key for the local tier.

Local-first, zero signup

The thing that kills most "agent memory" tools is friction: a signup wall between install and value. mammoth's default is local-only.

cargo install allsource-prime
curl -fsSL https://raw.githubusercontent.com/all-source-os/chronos/main/plugin/mammoth/install.sh | bash

That runs a fully durable store on your disk — the same AllSource Core engine (WAL with CRC32 + fsync, Snappy Parquet snapshots). Your memory survives reboots and never leaves your machine. No account.

Free tier remembers across sessions on this machine, no account. Cross-machine and team memory are the upgrade — add --sync-to + --api-key to ship memory to a hosted AllSource Core. A pull, not a gate.

Works in your agent

mammoth's core capability is one MCP server, so any MCP-capable agent speaks all 13 tools natively — Claude Code, Cursor, Cline, Windsurf, Codex, Gemini. One stanza, no per-agent shim:

{
  "mcpServers": {
    "prime": {
      "command": "allsource-prime",
      "args": ["--data-dir", "~/.prime/memory", "--auto-inject"]
    }
  }
}

Claude Code gets the richest setup — the MCP tools plus a mammoth-memory skill and /remember, /recall, /memory-status slash commands — via the plugin:

/plugin marketplace add all-source-os/chronos
/plugin install mammoth

Does it actually work?

Claims are cheap; mammoth ships a reproducible benchmark (tooling/mammoth-bench/) you can run yourself. It scores recall 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:

metric result
Recall hit@5 0.90 (baseline 0.83)
Recall hit@3 / MRR 0.87 / 0.783
Median tokens saved / recall 19 (986 total over 54 hits)
Recall latency p50 / p95 3.0ms / 3.6ms
Durability (write → restart → read) PASS

Honest note: on a smaller 18-memory set the edge over grep 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, and 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. That's the same trust move caveman made with "affects output tokens only."

Try it

cargo install allsource-prime
/plugin marketplace add all-source-os/chronos
/plugin install mammoth

Then work normally. A few sessions in, ask your agent why you made some past decision — and watch it remember.

Stop re-explaining your project. Give your agent a mammoth.

Immutable event sourcing with time-travel queries, 43 MCP tools, and x402 agent payments. Free tier — no credit card required.

Give your AI agents perfect memory

No credit card required. 10K events/month free.