Smart Memory + Knowledge — The Reason the AI Actually Remembers.

Generic AI assistants forget you the moment the chat closes. AeionClaw's Smart Memory extracts facts, preferences, context from every conversation and stores them as semantic embeddings in your tenant database. RAG-enabled retrieval surfaces relevant past context on every new turn. 10 knowledge categories for structured documentation. Per-tenant + per-persona partitioning prevents leakage. Same DB, same RBAC, same audit log.

Smart Memory — How It Works

Every conversation turn, AeionClaw can extract memory candidates — facts, preferences, context, and action items — each with a confidence score, an optional expiration for time-bounded information, and a partition that determines which persona is allowed to recall it later.

6 Memory Types — What Gets Stored

fact

Stable assertions about user or tenant ("CTO at Mercury Financial", "uses Stripe for billing")

preference

User choices ("prefers concise responses", "wants Slack notifications, not email")

decision

Something the user decided to do ("standardizing on Postgres", "moving Q3 launch to August")

relationship

The user's relationship to a customer or contact ("Carla is the primary buyer at Acme")

workflow

How the user likes to do things ("always reviews drafts before publish", "batches invoicing Fridays")

goal

What the user is working toward ("hit $2M ARR by year-end", "ship the events module")

10 Knowledge Categories — Structured Documentation

Beyond ad-hoc conversation memory, Aeion Workspace Knowledge Base provides 10 structured categories for institutional knowledge.

RAG Retrieval — How It Works in Production

Embed the user message

via the configured embedding model (e.g. Voyage AI voyage-3, Cohere embed-multilingual)

Vector search across memory + knowledge

using pgvector inside Aeion DB

Retrieve top-K most-similar entries

typically 5-15 per turn

Filter by partition + role + freshness

RBAC + expiration check

Rerank with Cohere Rerank

surface most-relevant for the current intent

Inject into LLM prompt

within token budget

Cite sources

AI response includes which memories/knowledge informed it

Audit + cost-track

embedding cost + retrieval latency logged

Memory Privacy + Per-Tenant Isolation

Memory is tenant-scoped. Personas are scoped. Roles are scoped. No leakage across boundaries.

Memory + knowledge questions

No hard cap. Typical customer at 100K active users: ~50-100K memory entries (most are deduplicated/expired). pgvector handles billions of vectors with proper indexing. Per-tenant cost caps prevent runaway growth.

Yes — GDPR data-portability API exports the user's full memory set. Per-user memory listing in the admin UI shows "what AeionClaw remembers about you."

Yes — `DELETE /memory/:id` endpoint. Full memory wipe via the user-erasure GDPR endpoint.

Memory = automatic extraction from conversations + user/persona-scoped. Knowledge = explicitly authored documentation + role-scoped + version-tracked. Both feed RAG. Both can be cited in responses.

Configurable per-tenant. Defaults to Cohere embed-v3 (multilingual). Voyage voyage-3 for cost-conscious. text-embedding-3-small for OpenAI-first stacks. Custom embedding models via BYOM.

RAG-quality benchmarks for our default config: top-5 retrieval recall ~92% on canonical benchmark (HotpotQA, FEVER). Cohere Rerank improves this to ~96%. Most retrieval misses are due to ambiguous query phrasing, not vector limitations.

Yes — every response includes a "sources" field listing retrieved memory + knowledge IDs. Admins can audit provenance.

Yes — that's the point. Memory is per-user, per-tenant, persistent in the DB. Lasts until explicitly deleted or expired (`expiresAt`).

An automated freshness check flags entries unreviewed for 180+ days. Owner notified. Stale memories are deranked in retrieval (lower score) and eventually auto-archived.

The AI that actually remembers.