io.github.hunta-ai/gather

Gather

Deterministic, attestable memory for AI coding agents.

Get started · Console · How it works · Compare

This repo is the open integration surface for Gather, a multi-tenant org-memory service (an MCP server plus a REST API) with server-enforced tenant isolation and attestable, curated writes. The memory engine is a hosted service; this repo holds the client integrations (starting with the Claude Code plugin) so your harnesses use that memory automatically, instead of hoping the model remembers to.

What's here

More integrations (Codex, Agent SDK) land here over time.

Quickstart (Claude Code)

Two lines, then a token:

/plugin marketplace add hunta-ai/gather
/plugin install gather@hunta

Mint an agent key in the console (scoped to propose + recall, never to write canon), then point the plugin at your tenant:

export GATHER_URL=https://mcp.hunta.ai   # or your own deployment on Estate (self-host)
export GATHER_TOKEN=<your-agent-key>

Recall and capture are live on your next session. Full walkthrough: hunta.ai/get-started.

Using another client (raw MCP)

Not on Claude Code, or prefer raw MCP? Point any MCP client at the server directly (see examples/mcp-config.json):

{
  "mcpServers": {
    "gather": {
      "url": "https://mcp.hunta.ai/mcp",
      "headers": { "Authorization": "Bearer <your-agent-key>" }
    }
  }
}

What the plugin does

• Auto-recall on every prompt. A UserPromptSubmit hook queries your memory and injects the top facts into the turn, so the agent starts with what it already knows.
• Re-inject after compaction. A SessionStart(compact) hook re-runs recall so injected context survives Claude Code's context resets.
• Loss-proof capture. PreCompact and SessionEnd hooks post a content-free breadcrumb to your staging inbox. The transcript is never uploaded.
• Curation skills. /gather:flush reviews pending candidates; /gather:status shows local recall hit-rate and latency.

The principle is harness-enforced invocation: tool availability is not tool use. Agents do not reliably recall context or write learnings back on their own, so the hooks make it happen by construction, on every turn, without depending on the model's judgement.

How memory works

Gather's differentiator is the write path: the writer never decides.

1. Capture lands candidates in a staging inbox. Candidates are invisible to recall.
2. Promotion runs the curated-write pipeline (extraction, dedup, bi-temporal supersession, provenance stamping). This is the only path into canon, and it is a human (or policy) decision.
3. Recall returns only sealed, attributable facts. An injected agent can propose noise; it can never seal a fact.

Always-on capture means you never lose a learning to a closed session; the staging wall means an unreviewed breadcrumb can never pollute the memory your agents read. Read more: the curation gate.

Fail-open by construction

A slow or unreachable memory server never blocks or degrades a turn.

• 8-second hook timeout.
• Every error path (timeout, 4xx/5xx, missing token, network failure) exits 0 with no output: inject nothing, continue.
• Automatic capture only ever adds candidates to staging, so canon is never touched on any error path.

Worst case for a down server is a turn with no injected context, never a stalled turn.

Privacy

Everything goes only to the GATHER_URL you configure. Recall sends the prompt text as a search query; capture sends a breadcrumb (event, cwd, session id, timestamp) and never the transcript. On the hosted service that endpoint is mcp.hunta.ai; on Estate it is your own deployment and nothing leaves your infrastructure. Full data-flow table: docs/plugin.md.

Why Gather

Honest comparisons, including where we're behind: hunta.ai/compare.

Community & support

• Issues and feature requests: GitHub Issues
• Security reports: see SECURITY.md
• Questions: hunta@agentmail.to

Contributing

Integrations and fixes welcome. See CONTRIBUTING.md and our Code of Conduct.

License

Apache-2.0.