Hunta Gather

Your agents propose. You approve. Gather remembers.

The steady, cumulative memory for your agent fleet. Agent keys can only propose, owner keys write directly, every tenant is isolated, and recall brings back what matters.

Hunta Gather

Agent memory with a receipt

Hunta Gather stores what your agents learn and recalls the current truth. Connect it to any agent harness, then run the check that sets it apart: verify your tenant's data is isolated, and get back an Ed25519-signed result you can check yourself.

token to first recall in five minutes

Get a token See the full flow →

/plugin marketplace add hunta-ai/gather
/plugin install gather@hunta
export GATHER_URL=https://mcp.hunta.ai
export GATHER_TOKEN=<your-agent-key>

# ~/.codex/config.toml
[mcp_servers.gather]
url = "https://mcp.hunta.ai/mcp"
 
[mcp_servers.gather.http_headers]
Authorization = "Bearer <YOUR_KEY>"

{ "mcpServers": { "gather": {
  "url": "https://mcp.hunta.ai/mcp",
  "headers": { "Authorization": "Bearer <YOUR_KEY>" } } } }
# hands-free install: point your agent at hunta.ai/connect/agent

{ "mcpServers": { "gather": {
  "url": "https://mcp.hunta.ai/mcp",
  "headers": { "Authorization": "Bearer <YOUR_KEY>" } } } }

# gather
curl -X POST https://mcp.hunta.ai/v1/memories -H "Authorization: Bearer $GATHER_TOKEN" -d '{"text":"Our design partner is Acme Corp."}'
# recall
curl -X POST https://mcp.hunta.ai/v1/memories/search -H "Authorization: Bearer $GATHER_TOKEN" -d '{"query":"design partner"}'
# verify (the beat a plain memory API cannot show)
curl -X POST https://mcp.hunta.ai/v1/admin/verify-isolation -H "Authorization: Bearer $GATHER_TOKEN"

POST /v1/admin/verify-isolation · signed resultISOLATED

The moats

Four reasons this holds up

Not leakableLive

Attestable isolation

Schema separation, row-level security, and per-tenant encryption, signed into a receipt you can reproduce and verify yourself.

Not poisonableLive

Governed write path

The agent doing the writing never decides what is kept. Agent keys can only propose; owner keys write directly.

ChronologyLive

Bi-temporal

What was true, and when you learned it. History and as-of queries are live.

RelevanceLive

Recall and relevance

Hybrid retrieval: keyword and vector fused, then reranked. Foraging, made exact.

Governed write path

Live

Agents can only propose; a person approves what becomes memory.

Every memory an agent proposes is reviewed before it enters shared memory: agent keys can only propose, owner keys write directly, and the agent that proposes a write never approves its own write. The raw input is never lost. Extraction proposes candidates, a deterministic gate screens them, and only then does a fact enter memory. Poisoned input cannot approve itself.

Recall and relevance

Recall

A question about a renewal date. Three memories are ranked; the newest true one rises to the top with its date and source.

The right memory, at the right moment.See how recall ranks memories →

Hunta Reckon

Memory that reasons, not just remembers.

Reckon reasons over your whole store offline, the facts, the guardrails, and the pending queue, then proposes how it should improve. It never writes on its own: every proposal goes through the same curated gate.

Explore ReckonRoadmap

Hunta Instinct

Some lessons should never be recalled. They should be reflexes.

When a mistake is costly enough, Gather promotes it into a reflex (we call this Hunta Instinct): a constant read before every action, not retrieved on request. Experience becomes recall, recall becomes instinct.

Explore Instinct

Hunta Instinct

An agent is about to clear a shared folder. A known-trap card intercepts it, and the agent runs the cleanup task instead.

Lessons become reflexes.See how Hunta Instinct works →

Everything taken in, kept clean, recalled by relevance.

The steady memory for your agent fleet. Create a key and connect, or read the full walkthrough.

Start free Get started