Roadmap

What Graphory has shipped recently, what we are actively working on, and where we are headed. Dates are quarters, not commitments. For commit-level history, see the changelog.

Last updated: 2026-05-22

Shipped recently

Highlights from the last 60 days. For the full list, see the changelog.

• Redis-backed public-token rate limiter - atomic Lua-script token buckets that survive deploys and fail closed, replacing the legacy per-process limiter.
• Public /health endpoint - 12-check baseline as structured JSON for status pages and uptime monitors, with module-level caching and lock-and-refresh.
• Structured /ingest rejection with register_business hint - missing-entity rejections now return an actionable next step instead of an opaque error.
• Per-user extraction queue with stuck-row watchdog - per-user fairness on the extraction pipeline plus auto-recovery of jobs that stop making progress.
• Stripe lazy-init for billing - billing module no longer fails to import when Stripe credentials are absent in dev or self-hosted runs.
• Credential auth-failure pass-through - upstream provider errors surface verbatim through connection_error notifications with system slug, error text, and fix hint.
• WorkOS-native vault naming + lazy migration - vault credential layout now follows WorkOS conventions, with on-read migration of legacy entries.
• Three new discovery MCP tools - get_latent_connections, get_stale_entities, and get_weekly_digest for proactive agent check-ins and Monday-morning briefings.
• Full-text search + AI query expansion - keyword search is substantially faster, and agents can now pass synonyms / known aliases as expanded_terms to broaden recall.
• New describe_schema MCP tool - agents can now introspect the live schema and trust configuration they are working against.
• Unified MCP endpoint at api.graphory.io/mcp - consolidated across all clients and docs.
• Per-org graph isolation hardened - removed all default-graph fallbacks; every request now resolves to an explicit org graph.
• B2-primary raw storage - collected source files now live in object storage with a local hot cache, lifting capacity concerns for large orgs.
• Custom-source auth improvements - dual-header authentication and better diagnostics for custom collectors.
• Privacy scrub of public docs - all examples, narrations, and benchmark reports now use synthetic placeholders.

In progress

Workstreams we are actively shipping against. Rough timing in parentheses - this is our best current estimate, not a commitment.

• Expanded MCP tool vocabulary for conversation memory (Q2 2026) - richer write-back primitives so agents can capture decisions, takeaways, and open questions as first-class graph entries, not just messages.
• Category-aware custom-connector authoring (Q2 2026) - hints and templates surfaced through configure_custom_collector so the user's AI (Claude, Cursor, ChatGPT) can infer source category, auth style, and extraction shape from a system's docs and wire up a working collector in one MCP turn.
• LongMemEval hybrid benchmark refinement (Q2 2026) - ongoing work on our hybrid evaluation harness to keep pace with the state of the art in long-horizon memory benchmarks.
• Universal extraction coverage (Q2 2026) - one deterministic extractor for any source, driven by the canonical schema and user corrections. Industry-agnostic by design.
• Schema + correction feedback loop (Q2-Q3 2026) - every user correction updates the extraction rules, so the system gets smarter with use across the whole customer base.
• Graph API search upgrades (Q2 2026) - improved relevance ranking plus 1-hop neighbor inclusion so keyword matches surface their immediate graph context by default.
• Action-governed write schemas (Q2 2026) - typed action catalogs for every write endpoint, with confidence gating, validation, and a reviewable queue for ambiguous suggestions.

On the horizon

Directions we are committing to in the next quarter or two, scope still firming up.

• Public MCP directory submission (Q2-Q3 2026) - listing Graphory in the public MCP server directory so users can discover and install with one click.
• Customer case studies (Q2-Q3 2026) - deep-dive write-ups of real Graphory deployments across different industries and workflows. See use cases for early patterns.
• SOC 2 roadmap for enterprise (Q3-Q4 2026) - formal compliance work to unlock larger enterprise deployments. Until then, security practices are documented in our privacy page.
• Scoped API keys and per-source visibility rules (Q3 2026) - finer-grained scopes (per-source visibility, key-level read-only) for the existing org-owner API key model.
• Richer agent discovery surfaces (Q3 2026) - surfacing latent connections, stale entities, and anomalies directly into the agents' weekly rhythm, not just on-demand.

Not on the roadmap (but we hear you)

Things the community asks about regularly that we are deliberately not building soon. We want to be honest about where we are spending our time so you can plan accordingly.

• Built-in LLM for synthesis. Graphory is the memory layer. We will not bundle a reasoning model, and there are no LLM API keys in the production system. You bring your own AI (Claude, ChatGPT, Gemini, Cursor, Windsurf, local models) and point it at the Graph API / MCP server. This keeps you in control of cost, privacy, and model choice.
• Vector embeddings for search. We do not embed or store vectors. Instead, agents pass expanded_terms at query time so the AI - which knows user intent better than any embedding - drives recall. This is a conscious design choice, not a gap.
• A chat UI that replaces your AI. The dashboard exists for sign-up, connection management, and auditing. The primary interface is your own AI via MCP, not our web app.

Related

• Changelog - canonical release history
• Use cases - patterns we see in the wild
• MCP Tools - current tool reference