Roadmap#

Public, public-issue-driven. Open a GitHub issue if you want to push something up the list.

Now (v1.0 — shipped)#

JavaScript / TypeScript SDK — Node and browser bindings
LlamaIndex integration (memgar-llamaindex)
AutoGen / CrewAI integration packages
Hosted gateway (memgar-gateway) — optional SaaS for non-Python stacks, runs the same engine via REST/gRPC
Production-trained transformer model distributed via signed release, opt-in via memgar.download_model()
Public benchmark — memgar vs Lakera Guard / NeMo Guardrails / Rebuff on PINT and our 464-sample corpus, results page on this site
Threat intel marketplace — community-contributed pattern packs with signature verification

Active learning loop — production traffic → hard negative mining → automatic retraining → canary deploy with quality gate
Dashboard UI — web dashboard for memgar operators (Grafana- driven or standalone Next.js)
VS Code extension — inline detection during agent development
Multi-language model support — explicit per-language pattern sets (DE, ES, FR, JA, ZH) and recall guarantees
Federated threat intel — secure cross-org pattern sharing
Cloudflare Workers / Edge runtime support
SOC 2 / GDPR documentation pack for enterprise deployment

Open questions we'd love help on:

Provenance-preserving memory writes — can we annotate every chunk in a vector store with its origin signature without breaking retrieval?
Behavioural baseline transfer — can baseline from one agent bootstrap a similar one without exposing PII?
Adversarial robustness of memory-context wrappers — what new envelopes do attackers find next?
Cross-modal poisoning — image / audio / video tools feeding into agent memory.

These were considered for v1.0 and explicitly postponed:

Pre-shipped transformer model artifact (intentional — the default training data overfits; see training docs)
Closed-source pattern packs
Auto-learned source trust
Telemetry phoning home