PLEXUS · V0.2
Don't build
in silos.
The AI nervous system for small business. Plexus plugs into the tools you already use, builds a bounded memory of how you work, and runs reusable skills that get sharper every time you use them.
26
D1 TABLES
Multi-tenant from day one
4
MODEL PROVIDERS
Workers AI · Anthropic · OpenAI · self-hosted
§
BOUNDED MEMORY
Hermes-style — bounded, curated, never bloats
FTS5
CROSS-SESSION RECALL
Every chat + skill output, instantly searchable
HOW IT WORKS
Four steps. The fourth one is what makes Plexus different.
Every other AI tool can do steps 1, 2, and 3. The loop closes on step 4 — Plexus writes structured notes to itself after every run, so the next time you (or anyone on your team) runs the same skill, it's already smarter about your business.
Your tools, your data
Connectors, email drops, dossiers. Plexus pulls structured data into D1 — summaries and facts, not raw embeddings of everything.
Reusable workflows
Automations (Closed Won → Sage invoice) or conversational helpers (draft customer reply). Same engine, two shapes.
Three memory ops
After every run, the cheap model proposes add / replace / remove ops against bounded stores. Per-skill memory and per-user profile, capped — never grows forever.
Frozen snapshot
Memory + your profile + dossier context load once per run as a cached prefix. Multi-step skills get ~10× cheaper on recurring runs.
WHAT WE SHIP TODAY
Six primitives. Every one wired to a real D1 table.
Bounded memory
MEMORY.md per skill (≤2200 chars) and USER.md per user (≤1375 chars). When full, the AI must replace or remove — not append. Forces librarian discipline instead of bloat.
Three explicit ops
add, replace, remove. The reflection model returns structured JSON, not free-form prose. Every change writes a memoryEvent audit row so you can trace exactly what was learned, from which run.
Model routing cascade
Per-skill model > per-thread override > workspace default > global. Workers AI for fast cheap defaults, Anthropic BYOK for high quality, OpenAI BYOK, or your self-hosted endpoint via Cloudflare Tunnel.
FTS5 cross-session recall
Every chat message and skill output is indexed at write time with AFTER INSERT/DELETE/UPDATE triggers. Plexus can find that thing you talked about last week without rehydrating the transcript.
Skills as portable bundles
Markdown body + structured execution chain (skillStep → skillRunStep). Automations fire on triggers; conversational skills walk a story of source → instruction → preview → confirm. Both visible in the Skills page.
Per-record authz
Every mutation calls a lib/authz.ts loader that checks workspace + scope + role. No 'trust the form' — the auth layer is centralised so cross-workspace leaks are structurally impossible.
THE MEMORY LOOP
Hermes-inspired. Bounded. Auditable. Yours.
Plexus borrows the bounded two-file model from Nous Research's Hermes Agent, then extends it for multi-tenant SaaS — per-workspace scoping, audit log, and a feedback hook so 👎 corrections flow back into memory as priority signals.
- §Rooted Coffee Co. Ltd resolves cleanly to C-2147; the 2024 'trading' record is a known duplicate.
- §DELIVERY-Q2 line is normal for this customer — last 4 runs all had one.
- §Robert prefers we lead with the rebate when quoting Tier 2.
- §Robert writes in short paragraphs, lowercase subject lines, no exclamation points.
- §Signs off as 'R', not 'Robert'. Defaults to Net 30 unless hospitality customer (then Net 14).
- §Pinned reminder: surface shipping mentions on Monday.
EVERY CHANGE LOGGED
memoryEvent(workspaceId, target, targetId, op, content, sourceRunId, sourceUserId, createdAt)
Every add / replace / remove writes an audit row tied to the source run. The MemoryFile panel in the Skill page reads these to show the design's “Promoted from R-014 feedback” provenance — real, queryable, not a mock.
MODEL ROUTING
One skill spec. Four providers. Cascading override.
Plexus resolves the model on every call via a four-level cascade: per-skill override → per-thread override → workspace chat default → global default. Swap one tier, the rest still work.
Cloudflare Workers AI
DEFAULT@cf/meta/llama-3.1-8b · @cf/qwen/qwen-2.5-coder · …
Lives on the same edge as Plexus. No round-trip outside Cloudflare. Cheap default for fast skills + reflection.
Anthropic Claude
BYOKclaude-sonnet · claude-opus · claude-haiku
Bring your own key. Best for high-stakes runs — content with brand voice, structured reasoning, long-form drafts. Prompt caching wired.
OpenAI
BYOKgpt-4o · gpt-4o-mini · o1
Same BYOK pattern as Anthropic. Slot in for skills that need OpenAI-specific behaviour or where the customer already has a key.
Your own endpoint
SELF-HOSTEDOllama · vLLM · LM Studio · anything OpenAI-compatible
Expose via Cloudflare Tunnel, paste the URL into Admin → AI. Plexus uses your hardware. Air-gapped, private, free per-token.
SKILLS
Two shapes. One engine. Both grounded.
A skill fires every time something happens in a connected system — Salesforce opportunity Closed Won, calendar event 30 min out, new email matching a label.
Closed Won → match customer in Sage → build invoice → email + Stripe link
- ·Preview mode for the first N runs (you approve each)
- ·Auto-promotes to Live after N clean runs
- ·Full audit trail per run with per-step rationale
You walk the skill through a story: source → instruction → preview → confirm → next source → task → destination. Plexus shows what it found at each step so you can correct before it commits.
Reply to Hannah — read Gmail thread → pull tier sheet → check contract → match her tone → draft → save to Gmail draft
- ·Five preview kinds: email-thread / table / list / files / callout
- ·Grounded-in chips show every source the draft used
- ·Save the whole flow as a new skill in one click
The killer move:after any conversational run that felt useful, Plexus prompts “save this as a skill?” — and the AI drafts the skill recipe from the conversation. Every successful ad-hoc task becomes a reusable asset.
TRY IT
One click. Real skill. Live in Preview mode.
Sign in, spawn the demo skill (Closed Won → Sage invoice), see the full AutomationView story — blueprint, rules, memory, run history, per-step rationale — wired to real D1 rows in your workspace.