The organizational memory layer for the AI enterprise.
mind0 gives your agent fleet one shared, scope-controlled memory — user, agent, session, org. Agents plan against it, search it, and write to it; your team watches that memory form in real time.
17 API endpoints · 4 memory scopes · streaming tool-chains · live cost telemetry
Every layer below is shipped and running — the same stack that serves the console.
- Encore
- PostgreSQL
- pgvector
- GLM-5.2
- Langfuse
- Next.js
- TypeScript
- TanStack Query
- Radix UI
- Tailwind CSS
- Vercel
The console, at a glance
Every surface below ships today. The public page and the operator console share one navigation source — what you see here is what your team operates.
Memory
- TimelineLive memory event feed
- RecallSemantic search over memories
- ReflectSynthesize recent memories
- RetainWrite a durable fact
Agents
- RegistryThe registered agent fleet
- ChatTalk to an agent, tools visible
- IngestUpload data into shared memory
Organization
- SystemPilot KPIs and backend health
- WorkflowLive memory pipeline map
- GraphExtracted entities and relations
- MembersOrg roster and role management
Agents forget. Every session starts from zero, every handoff loses context, and every token spent re-learning is money burned. mind0 makes memory an organizational asset — scoped, searchable, and self-improving.
How memory forms
Four operations, one shared store. Every step below is a deployed pipeline stage, not a roadmap item.
- 01
Retain
Agents and people write durable facts into shared memory. Every write is scoped — user, agent, session, or org — and embedded on arrival.
POST /memory/retainscope checkembedpersist
- 02
Extract
An async knowledge-graph pass lifts entities and relations out of every memory, building the organization's graph as a side effect of normal work.
pub/subentity + relation extractiongraph
- 03
Recall
Semantic recall filtered by the scope visibility chain. Conflicting facts supersede at ≥0.9 similarity, so agents plan against the current truth — not the whole history.
POST /memory/recallvector searchscope filtersupersede
- 04
Reflect
Reflection synthesizes recent memories into higher-order insights and retains them back — the memory improves itself between conversations.
synthesizeinsightretain
Built for the enterprise floor
Multi-tenant boundaries, deterministic scope visibility, and cost you can put in front of a CFO — verifiable in the console, not a compliance slide.
- Scopes
- Every read walks the visibility chain — an agent sees its own memory, its session's, its user's, and the org's, never a sibling tenant's.useragentsessionorg
- Conflicts
- ≥0.9Contradicting facts supersede automatically — agents plan against the current truth, not the whole history.
- Tenancy
- JWT auth with per-request tenant identity; org data never crosses tenants
- Observability
- Exact token usage and cost per answer, per agent, per model — in-app
- Rate limits
- Per-tenant token buckets on chat and retain paths
- Tracing
- Optional Langfuse traces for every generation and reflection
- Storage
- pgvector semantic index plus an extracted entity-relation graph
- Interface
- 17A typed TypeScript client covers every console surface — no undocumented calls.
Numbers from the floor
Every figure below is a shipped constant or a live production measurement — nothing is a projection.
documented endpoints
The whole console runs on the same typed client you get.
verify: docs → typed client
memory scopes
user · agent · session · org — enforced on every read path.
verify: console → /org
supersede similarity
≥Contradictions auto-resolve at this threshold; history is kept.
verify: console → /timeline
measured cost per answer
$A real production answer: 1,752 tokens on glm-5.2 — not an estimate.
verify: console → /system → cost & usage
Watch memory form in real time.
Create an account, open the console, and see recall, supersede, and reflection run live.
point — a retained factline — an extracted relation
