Persistent memory for AI agents

CogniMemo is a note-taker for agents: your agent does the work and tells CogniMemo what it learned; next time, CogniMemo hands that knowledge back. It owns durable, typed storage and fast recall — it never executes or drives your agent.

Core operations

Retain, recall, reflect

Three operations cover the full memory lifecycle for production agents.

Retain

Store a memory. Embeddings, entities, and links are computed on ingest.

Text · files · URLs

Recall

Retrieve memories via hybrid semantic, keyword, graph, and temporal search — fused and reranked.

Typed filters

Reflect

Generate a synthesized, disposition-aware answer from memories and consolidated observations.

Synthesis

Memory types

Eight typed memory blocks

Three types are produced automatically; five more are typed blocks you write directly. Recall can filter to any subset — default recall returns them all.

worldexperienceobservationprocedurereasoningpreferencecorrectionprofile

Architecture

Gateway, engine, and dashboard

Each organization gets an isolated Postgres schema. API keys are scoped to an org and can be restricted by operation type.

Gateway

API key auth, rate limiting, quota enforcement, Stripe billing

Managed cloud

Engine

Per-org Postgres schemas, vector embeddings, entity graphs

Self-hostable

Web

Next.js dashboard for managing keys, viewing usage, and billing

Console

Lifecycle

Memories that stay relevant

Memories decay through states — New → Active → Expiring → Forgotten — so stale knowledge fades while frequently-used memories stay strong.

  • Corrections and explicit updates reset relevant memories
  • Entity graphs connect facts across sessions
  • Strict isolation — API keys never leak across projects