Memory as infrastructure
for AI.
An entity-specific associative network for AI. Vector search finds the entry points; the graph pulls in everything associated.
A memory graph, not a folder of notes.
Most AI memory is flat text with keyword lookup. OAK.memory stores each fact as a node — with a semantic embedding for similarity and structural edges for association — and retrieves them the way a mind does.
recall
A query enters the entity and traverses the graph — hop by hop — surfacing what plain similarity would never reach.
insert
A new fact is embedded and linked into the network — instantly associated with everything it relates to.
update
A memory's value is rewritten in place — the node stays, its meaning moves. No duplicates, no drift.
merge
Duplicate memories collapse into one canonical node — edges rewired automatically, nothing lost.
delete
A memory is pruned and its edges retract — forgotten cleanly, with no dangling links left behind.
oak-memory plugin · recall · insert · update · merge · delete — across every session
Edges carry no embedding.
A link is a pure structural connection, so what it means is interpreted at query time rather than frozen when it was created.
import { Memory } from "@openaikits/memory"const memory = await Memory.initialize({ db, embeddings })// vector entry → graph traversalconst ctx = await memory.recall("what does the user prefer?")await memory.remember("prefers dark, ships fast")▍Hybrid vector + graph
One unified store: embeddings for discovery, edges for association. Two-phase recall in a single call.
Entity-specific networks
Every user, persona, or agent keeps its own associative network. Personalization lives in the graph, not in the model weights.
Local embeddings
Run embeddings on-device via Ollama. No API key, no per-call cost, no data leaving the machine.
Framework-agnostic
Pluggable database and model adapters. Drops into LangChain, custom stacks, or any LLM framework.
One engine, many surfaces
Add memory to anything.
The engine knows nothing about any single app — it's infrastructure. Install the package, or drop it into Claude Code as a plugin and it recalls and stores facts about you across every project and session.
Where memory is going.
OAK.memory is the substrate. On top of it we're building the layer that lets machines — and people — truly remember.
Cloud Memory for Enterprise
Managed, secure memory infrastructure for teams and their agents, at scale.
Memory-native Assistant
A personal assistant that actually remembers you — across every session and device.
Human Memory Cloning
Capture and reconstruct a person's associative memory graph.
Action Foundation Model
A foundation model with memory-manipulation-optimised embeddings and native actions.
Why we build this.
Our Vision
Memory belongs in the stack as infrastructure — the way inference and model providers already are. General-purpose and domain-independent, not re-invented inside every application.
Our Mission
To deliver state-of-the-art AI capabilities with unmatched convenience and cost — so anyone can give their AI a real, persistent memory without the complexity.
What we optimize for.
Innovation
We turn the latest AI research — associative memory, hybrid retrieval — into practical, drop-in tools.
Accessibility
A few lines of code, local by default. Sophisticated AI infrastructure without the operational overhead.
Efficiency
Local embeddings and a lean retrieval path keep cost and latency low while performance stays high.
Let's talk.
Building with OAK.memory, or want it in your stack? Send a note.
Independently built and maintained.
