Agentic SOC memory — private beta

Memory for Agentic SOCs

The operating system for threat recall. A persistent, queryable memory layer for AI agents, agentic SOC platforms, and the analysts who run them.

Request access Read the origin story → Read the docs
threatrecall — agent recall query
> What do we know about APT29 lateral movement via PowerShell?
AGENT MEMORY 2 weeks ago · jsmith

APT29 uses PowerShell remoting for lateral movement targeting Exchange servers. Linked to CVE-2024-1234. Observed across 3 MSSP client environments.

SYNTHESIS

APT29 favors PowerShell remoting via scheduled tasks for lateral movement. Prefers targeting Exchange and AAD Connect servers. Recommend hunting T1059.001. Confidence: HIGH — 3 evidence chains.

AGENT langchain-agent · session #441

Query completed. 2 investigations recalled, 1 knowledge graph branch retrieved. This session now carries APT29 context — no re-investigation needed.

Agentic memory · Structured CTI · Knowledge graph
The problem

AI security agents start from zero every session. Analysts lose everything to turnover. CTI tools are built for browsing, not recalling.

01

Agents forget

Every AI agent session is a fresh start. LangChain, AutoGen, CrewAI — none of them persist what they learned across sessions. The same investigation gets re-done by the same agent, every time.

02

Analysts lose institutional memory

Years of threat intelligence — actor behaviors, victim profiles, infrastructure patterns — lives in analysts' heads, Slack threads, and personal notebooks. Turnover erases it all, permanently.

03

CTI tools are for humans, not agents

Existing threat intel platforms assume a human at a keyboard browsing dashboards. They weren't designed for agents to query, retrieve, and synthesize. You're left building one-off integrations that break on every API change.

Built for agents AND analysts

API-first. LangChain shipped. MSSP-ready from day one.

ThreatRecall integrates with your agentic stack — not a human dashboard bolted onto an API.

1

Agent writes to memory

Agents POST structured threat observations via REST. ThreatRecall extracts entities, resolves aliases, and links to the knowledge graph automatically — no manual tagging.

// LangChain agent calls:
recall.write(
"APT29 using PowerShell remoting for lateral movement",
agent_id="cobalt-strike-v2",
session="soc-prod-01"
)

// ThreatRecall extracts & links:
Actor: APT29 → TTP: T1059.001
CVE: 2024-1234 → IOC: 192.168.x.x
Knowledge graph: 3 edges created
2

Knowledge graph grows

Every write feeds the graph. Actor relationships, campaign attribution, IOC infrastructure chains — all connected. Multi-tenant isolation means MSSPs can serve N clients with completely separate memory stores.

APT29
uses
T1059.001
targets
CVE-2024-1234
leads to
Campaign #12
3

Any agent retrieves

Natural language recall. Vector + keyword + structured routing. Agents and analysts both query the same memory layer — same trust, same recall quality.

// Agent or analyst queries:
recall.search(
"What do we know about APT29 lateral movement?"
)

// Returns: structured response + graph branch + evidence chain
4

Integrations

LangChain adapter shipped. MCP server planned. OpenCTI, CrewAI, AutoGen on the roadmap. Your agentic platform — ThreatRecall memory layer underneath.

LangChain MCP (planned) OpenCTI (roadmap) CrewAI (roadmap) AutoGen (roadmap)
What ThreatRecall is

Mem0 gave AI agents personal memory. ThreatRecall gives agentic SOCs collective threat memory.

Purpose-built for AI agents, not bolted on. Three pillars that generic agent memory platforms can't offer.

Structured Recall

Actors, CVEs, IOCs, TTPs stored as typed entities — not embeddings against raw text. Agents query by type, confidence, TLP, and temporal bounds. No more vector-similarity soup.

vs. Mem0 / generic memory — unstructured text, no entity typing

Knowledge Graph

Causal relationships, alias resolution, campaign attribution. Actor → CVE → IOC → Campaign chains as first-class citizens. Every node carries confidence scores and evidence provenance.

vs. flat key-value memory — no graph traversal, no relationship semantics

Intent-aware Retrieval

Keyword + vector + structured CTI routing. Agents pass natural language; ThreatRecall routes to actor search, CVE lookup, IOC match, or graph traversal — whichever fits. STIX 2.1 compatible output.

vs. single-vector search — no structured filtering, no CTI semantics

Security-grade by default

TLP propagation, STIX 2.1 compatibility, evidence chain-of-custody, FedRAMP Moderate baseline. Multi-tenant isolation. OCSF-compliant audit logs. This is what generic agent memory tools cannot offer — not a bolt-on, it's the foundation.

vs. Mem0, Letta, etc. — no compliance, no TLP, no STIX
What generic agent memory can't offer

Built for SOC workflows from the ground up.

ThreatRecall was designed around how security operations actually work — not adapted from general-purpose agent memory.

TLP Propagation

Traffic Light Protocol labels on every memory entry. Agents and analysts respect TLP boundaries — AMBER data stays with AMBER-clearance users, not leaked to the general memory pool.

STIX 2.1 Compatibility

Structured data in, STIX objects out. Actor, Malware, Indicator, Vulnerability — typed entities export as standard STIX 2.1 bundles. Integrates with MISP, OpenCTI, and any STIX consumer.

FedRAMP Moderate Baseline

RBAC, audit logging, FIPS 140-2 encryption at rest and in transit. ATO-ready for U.S. federal and defense-adjacent environments. FedRAMP package available upon request.

Evidence Chain-of-Custody

Every IOC, CVE, and actor attribution carries provenance. Source, confidence score, linked evidence records. Audit trail that satisfies legal and compliance review — not just analyst convenience.

Knowledge Graph

Threat relationships as first-class data.

Every ingested observation grows the graph. Actor aliases resolve. Campaign chains link. Evidence trails propagate.

exploits generates part of attributed to ACTOR APT29 conf: HIGH CVE 2024-1234 CVSS: 9.1 IOC 192.0.2.44 TLP: AMBER CAMPAIGN SolarStrike-12 3 incidents T1059.001 alias resolved ✓
Actor CVE IOC Campaign Directed edges animate in real time as agents write memory
Open-source or managed — your call

Two ways to run ThreatRecall.

Same memory engine. You pick the operating model. No lock-in either way.

MIT License

ZettelForge

Self-host the memory engine. Free, MIT-licensed, authored by Patrick George Roland II. Typed memory model, blended retrieval, FastAPI interface. You run the infra.

  • LanceDB + SQLite — no managed DB required
  • Ollama-compatible — your models, your hardware
  • LangChain + CrewAI adapters included
  • Full source — audit it, fork it, own it
Managed cloud

ThreatRecall Cloud

Managed, multi-tenant, FedRAMP-aligned. We run the infra so your team focuses on threat intelligence — not ops. Enterprise-grade from day one.

  • Multi-tenant RLS — tenant isolation at DB layer
  • LLM hosted — we run OpenAI, you bring context
  • FedRAMP Moderate baseline + STIX 2.1
  • SLA-backed + audit logs out of the box
Modeled on the WordPress.com vs .org model — same engine, two operating modes. See full comparison →
Who it's for

Agentic SOC platform builders

You're building an AI-native SOC product. You need a memory layer your agents can query and write to — with STIX compatibility, TLP controls, and audit logging built in.

MSSP operators

Multi-tenant threat memory across N client environments. Every analyst on every client gets the full recall of every engagement — no more siloed knowledge.

Threat intel teams (AI-augmented)

Your analysts are running AI agents. ThreatRecall gives those agents memory that persists, that carries context between sessions, that makes every AI-assisted investigation better than the last.

Solo researchers and red teamers

Your findings disappear after every engagement. ThreatRecall gives you a searchable threat memory — not a messy notes folder — so 6 months later you can ask "have I seen this before?" and actually get an answer.

From the Founder

"I'm Patrick Roland — Navy veteran, former MSSP director. Every SOC I've worked with loses investigations to agent session resets and analyst turnover. ThreatRecall is the memory layer I wanted from day one — for the AI agents we're all building and the analysts who need them to actually work."

Pricing

Three tiers. One memory layer.

Individual Researcher ($49/mo) and Pro ($199/mo) — live and available. Enterprise — contact sales.

Individual Researcher
$49/mo

Solo researchers and red teamers. Persistent threat memory across engagements. Natural language recall. Searchable history you can actually query.

  • Knowledge graph
  • REST API access
  • 1 workspace
  • STIX export
Enterprise
Contact us

MSSP operators and agentic SOC platform builders. Multi-tenant isolation. FedRAMP Moderate baseline. Dedicated infrastructure. Custom SLAs.

  • Everything in Pro
  • Multi-tenant (N clients)
  • FedRAMP Moderate baseline
  • Dedicated infra
  • Custom SLA
  • OCSF audit export

Start your pilot.

ThreatRecall is live. Individual Researcher ($49/mo) and Pro ($199/mo) plans are available. Enterprise — contact sales.

Individual Researcher Pro Enterprise