Five operations and an entity graph: store, search, remember, erase, status, graph search. An agent reaches them over MCP, REST or an SDK.
Memory works in two tiers. New material lands in working memory; a background pass consolidates it into long-term facts and a graph of entities and relations — the handoff the design is modelled on.
Every answer carries its sources. When nothing in memory grounds an answer, the service says so instead of inventing one.
In regulated work the question is not whether a memory service is good. It is whether the data may leave the building at all.
For a law firm confidentiality is a professional duty; for a clinic and a bank it is a supervisor's requirement. A product that keeps your agent's memory on someone else's infrastructure ends that conversation before any feature matters.
Every row carries its tenant. Row-level security with FORCE ROW LEVEL SECURITY applies the policy even to the table owner, so a query without a tenant context returns nothing — it cannot return another tenant's rows. The boundary is not application code that a future change might bypass.
A bleed test attempts a cross-tenant read on every build. If it ever succeeds, the build fails and nothing ships.
A signed file, verified by an Ed25519 key built into the image. No activation server. Billing cannot be an exception to "your data stays here".
When a licence lapses, reading and search keep working and only new writes stop. An agent that lost its knowledge over a late invoice does not come back.
Most memory products can tell you what they stored. This one answers for it: every source that contributed to a fact, and every step its confidence took.
A fact becomes a theorem only through confirmations from independent sources — different documents, agents or reviewers — until it reaches 0.9 confidence; an agent repeating itself, or verifying its own statement, counts once. An agent can ask for theorems only. What nobody vouched for stays flagged for verification; a disputed fact is not promoted while the dispute is open, and a correction supersedes the old value with its history kept.
GET /memory/facts/{id}/why
{
"statement": "Renewal notice goes to the compliance officer",
"claim_type": "theorem", // independent sources, confidence ≥ 0.9
"confidence": 0.9,
"epistemic": "confirmed",
"reinforce_cnt": 3,
"distinct_sources": 4,
"reinforcements": [
{ "reinforced_at": "2026-05-19T10:02:11Z", "delta_confidence": 0.05,
"source_url": "minutes/board-2026-05.md" },
{ "reinforced_at": "2026-03-04T08:40:55Z", "delta_confidence": 0.05,
"source_url": "contracts/2026/addendum-7.pdf" },
{ "reinforced_at": "2026-02-11T15:20:03Z", "delta_confidence": 0.05,
"source_url": "policies/notices-v3.md" }
]
}Confidence is never raised by a fact merely appearing in search results — being shown is use, not corroboration.
These describe the shape of the work the product is built for. Haw0k-Mem0 is new and has no public references yet — none of the below is a named customer.
One deployment, separate memory per customer, isolation guaranteed by the database rather than by your application code. Priced per tenant.
Each matter in its own namespace. The agent recalls case facts across sessions, and the provenance trail answers where each one came from.
Credit scoring is a high-risk use under Annex III of the EU AI Act; from 2 December 2027 deployers must keep its logs for at least six months. Every write, correction, confirmation and deletion goes into an append-only audit log in your own database — time, tenant, agent, key and ids, never the content — kept for good or as long as you set, never under six months.
Departments isolated from one another. Secrets and contact details are masked before anything is written, and a namespace can be erased on request.
Direct to organisations, per year
Evaluate it on your own data
One organisation, a few agents
Several units, or a growing platform
No limits, contractual guarantees
| Trial | Team | Business | Enterprise | |
|---|---|---|---|---|
| Tenants | 3 | 1 | 10 | ∞ |
| Agents | 5 | 3 | 20 | ∞ |
| Entity graph | ||||
| Re-ranking | ||||
| Verification contour | ||||
| Support | — | Next business day | SLA | |
| EU AI Act pack | On request |
Every paid tier includes the entity graph; building it needs an LLM endpoint you configure, local or remote. Business adds precision on large volumes and control over what the agent was allowed to learn.
Priced per tenant: €360 up to 25, €240 for 26–100, €150 for 101–500. Minimum contract €6,000 per year. White-label rights are included from 26 tenants.
Tell us which of your agents needs memory. We issue a signed 30-day trial licence in your organisation's name and grant access to the image.
The trial covers three tenants and five agents — enough to watch the isolation hold. Nothing phones home at any point, during the trial or after it.