Media kit
Status: positioning + one-liners + sizing for a landing page, a PR pitch, or a journalist. Author-faithful to the product (not an external analyst’s endorsement). Version-grounded: every technical claim maps to a shipped release in the proof map.
Name / one-liner
- Product: Brain Server
- One-line (technical): “A local-first decision and memory substrate for AI agents — deterministic retrieval, human-gated state promotion, tamper-evident provenance, and structured decision traces.”
- One-line (primary positioning): “Governed decision and memory substrate for AI agents — deterministic local recall, human-gated permanent state, and tamper-evident audit, with structured decision traces.”
- One-line (buyer): “Agent memory and decisions you can verify, budget, and delete on request — no LLM per query, no data egress, no vendor lock-in.”
- Three-word elevator: “Verifiable agent decisions.”
- One-line (contact-center / BPO support): “Agent-assist memory and decision substrate that recalls past resolutions and policy, stays on-prem, and is yours to audit and erase — no per-query LLM, no vendor lock-in.”
Positioning statement
For teams building AI agents that must hold memory and make structured decisions responsibly, Brain Server is a self-hosted substrate that makes both recall and the decisions that depend on it deterministic, human-gated, and tamper-evident — unlike cloud memory services that charge per query and keep user data in a third-party datacenter.
Because it runs on the operator’s own infrastructure with no LLM in the hot recall path, it delivers zero per-query cost, zero data egress, and an audit trail a reviewer can verify live.
Who it’s for
The same engine serves several audiences; see Who it’s for — target audiences for the full map (each marked shipped vs. roadmap).
- AI-agent builders & OpenClaw users — deterministic memory, zero token cost, in the memory slot.
- BPOs & multi-client contact-center operators — the v2.0 “Cortex” roadmap
is explicitly call-center intelligence (multi-team tenancy, ticket-pattern
resolution). The controls they need are shipped today (per-domain
scoping, per-tenant audit, DSAR, PII containment, human-gated writes);
multi-client tenancy on one shared backend is the documented v2.0 piece
(true storage isolation is the separate
BRAIN_MULTI_DBmode, not the default shim). - In-house contact & support centers — agent-assist memory that recalls past resolutions and policy, supervised and audited, without fabricating answers (calibrated abstention + span verification).
- Regulated enterprises (finance, healthcare, legal, government) — memory that stays on-prem, is auditable to a chain, honors DSAR, and is explainable.
- Edge / field / air-gapped deployments — a single self-hosted runtime on 4 GB ARM.
- Delivery partners (SIs, MSPs, consultants) — a deployable, auditable
memory layer with procurement-grade evidence (
RFP_RESPONSE_KIT.md).
The three pillars (press-ready)
- Recall that never has to think — deterministic, reference-faithful retrieval (bi-temporal KG, submodular packing, PPR graph leg, hub dampening, calibrated abstention). No LLM decides, no token is spent.
- A write gate, not a write path — memory is proposed and promoted only on human approval; an injection screen quarantines adversarial input.
- A chain, not a log — every decision lands in a tamper-evident SHA-256 chain; DSARs produce chain-verifiable deletion certificates; an OWASP 2026 control matrix states every control as shipped or owned ceiling.
Brain vs. the field (sizing, with honest ceilings)
| Brain Server | Mem0-class (framework memory) | LangGraph-class (agent framework) | Plain RAG | |
|---|---|---|---|---|
| Per-query cost | $0 | LLM/embedding API | LLM/embedding API | LLM/embedding API |
| Where memory lives | Your device | vendor/cloud | vendor/cloud | your infra |
| Recall determinism | Yes | no | no | partial |
| Human write gate | Default | optional | no | no |
| Tamper-evident audit | Yes (hash chain) | no | no | no |
| DSAR deletion cert | Yes | partial | no | no |
| Standard wire | UMP L3 + open HTTP + MCP | proprietary/framework-bound | framework-bound | none |
| Zero LLM in loop | Yes | no | no | no |
Honest ceilings we don’t claim (each owned + versioned): multi-team tenancy (v2.0), per-tenant limits (v2.1), pricing/licensing (v2.2). Retrieval is deterministic, not SOTA-generative; multi-hop graph quality is corpus-bound; abstention is heuristic, not learned.
Headline stats (verify in the proof map)
- UMP 1.0 — 13/13 reference-suite checks (
@universalmemoryprotocol/core, CI re-run): L3 signed on a keyed instance, L2 hash-only without a key. $0per query — no LLM/embedding API in recall or writes.- Small-device capable — runs on a 4 GB ARM device (Jetson Nano / RPi 5); no power-draw figure is claimed (none measured).
{"ok":true}in one command —/audit/verifyproves the chain intact.- OWASP 2026 matrix — 100% control coverage (shipped or owned ceiling).
Press contact / ask
For a reviewer: run the 3-minute reproduce.md walk-
through to verify every security claim live against a throwaway instance —
“trust us” becomes “verify it.” For a journalist: the honest-ceiling post
(blog/07-honest-ceiling.md) is the story — a
memory store that tells you its limits.
Logos / naming notes
Name has no built-in icon yet (operator step). The wordmark is “Brain Server”;
the CLI/product family is brain / brain-server / mcp. Repository:
markfietje/brain-server.
Author / contact
Maintained by Mark Fietje:
- LinkedIn: linkedin.com/in/markfietje