Memory EngineShipped
Rust engine in Cognee Cloud
The Rust implementation of the cognee pipeline runs production workloads on the cloud and in dedicated customer deployments.
Roadmap
Where cognee is heading.
Cognee is built in the open, so the plan should be too. This page tracks what we shipped, what the team is working on this quarter, and the directions we are committing to next. We update it every quarter.
Items under Now and Next are commitments. Later and Horizon describe where we are investing design and research time; they are directions, not delivery dates.
SIX AREAS OF WORK
Memory Engine
The open-source core: ingestion, graph construction, recall, and the Rust engine.
Cloud
Cognee Cloud, dedicated deployments, and the infrastructure under both.
Company Brain
Memory for teams: coding agents, internal tools, and the knowledge that connects them.
Research
Where we push on the science of memory, evaluations, and self-improvement.
Ecosystem
Integrations, SDKs, MCP, and the community around cognee.
Platform
Security, reliability, and the operational layer every deployment inherits.
Where we disagree with the field
Most of what follows is a consequence of six positions. They are not popular, and we hold them anyway.
- 01
Memory should forget.
Append-only memory is a liability dressed up as a feature. Facts should expire, decay, and be deleted with proof that they are gone.
- 02
Inference is memory too.
A system that concludes things about you and keeps them hidden is worse than one that writes them down where you can read, dispute, and delete them.
- 03
Memory belongs to the person, then the organization, then the vendor.
In that order, and enforced by the engine rather than by a policy document.
- 04
No lock-in, including ours.
We will ship the exporter to our competitors before they ship one to us.
- 05
Memory has to work without an LLM.
If the pipeline falls over when the model API does, it is not memory. It is a cache.
- 06
A memory you cannot audit is a rumor.
Every fact carries its source, its time, and the reasoning that produced it, or it does not get stored.
Recently shipped
The foundations the rest of this page builds on.
CloudShipped
Stateless tenant architecture
Cognee Cloud moved to a per-tenant, stateless design with sharded orchestration and cold starts measured in seconds.
ResearchShipped
Self-improving memory loop
Memory that learns from how it is queried: feedback on recall shapes what gets strengthened, merged, or retired.
EcosystemShipped
Cognee MCP server
Persistent memory for Claude, Cursor, Codex, and every other MCP-aware client, backed by the same graph as the SDK.
Company BrainShipped
Company Brain
One memory layer across a team's documents, conversations, and code, available to people and agents alike.
EcosystemShipped
Rust SDK
A native Rust SDK alongside Python, for teams that need the engine embedded in their own services.
Memory EngineShipped
Cognee Edge
The Rust engine running on phones, wearables, and vehicles with no server in the loop. Memory forms where the data is and syncs when it can.
Now
Q4 2026What the team is building this quarter. Most of this is already in open pull requests or behind flags on the cloud.
Memory EngineIn progress
Temporal memory
Timestamp-anchored recall as a first-class path: ask what was true at a point in time, through the SDK, CLI, and REST API.
Memory EngineIn progress
Recall routing v2
A smarter router that picks the right retrieval strategy per question, with new temporal, code, and broad-search routes and an accuracy benchmark behind it.
Memory EngineIn progress
Contradiction resolution
When new information conflicts with what memory already holds, cognee resolves it explicitly instead of storing both and hoping.
Memory EngineIn progress
Horizontal scaling
Cognify runs across CPU cores and multiple API workers, so one heavy ingestion job can no longer stall everything else.
ResearchIn progress
Memory without LLMs
Replacing LLM calls in the pipeline with purpose-built models where they are not needed, cutting cost and latency while holding benchmark scores.
CloudIn progress
Scale-to-zero tenants
Idle cloud tenants scale to zero and wake on first request, so the free tier costs nothing when it is not being used.
CloudIn progress
Multi-cloud control plane
A cloud-agnostic tenant stack with a full reference deployment on Google Cloud, so customer data can stay in the cloud they already run.
CloudIn progress
Bring your own cloud
An installable bundle of the cognee stack for customer-owned clusters, with a preflight checker and a defined upgrade and support model.
Company BrainIn progress
Coding agent to Company Brain
Memory captured by coding agents flows into the team's shared brain, and context from the brain flows back into the next session.
Company BrainIn progress
Memory-aware code review
A reviewer that remembers prior decisions, past incidents, and the reasoning behind the codebase, running first on our own pull requests.
EcosystemIn progress
Native integrations in the cloud
Twenty integrations that bring cognee into the tools people already work in rather than beside them.
EcosystemIn progress
Memory in desktop AI apps
Cognee inside Claude Desktop and ChatGPT Desktop through MCP, so personal memory follows you across assistants.
PlatformIn progress
Security audit and CVE pipeline
Advisory triage, hardened server defaults, connector review, and automated CVE scanning in CI across the Python and Rust codebases.
ResearchIn progress
Continuous evaluations
LoCoMo and task-specific benchmarks run automatically on every change, producing reproducible scores for speed and quality.
EcosystemIn progress
Open community backlog
Working through the open pull requests and issues across cognee, cognee-community, and cognee-integrations with a dedicated maintainer.
Next
H1 2027Committed work that depends on what ships now. Scope is set; sequencing may change.
Memory EnginePlanned
Forgetting on purpose
Every fact carries an expiry and a decay curve. Deleting produces a signed certificate of what was removed, from where, and when. The opposite of append-only.
Memory EnginePlanned
Inferred memory
Cognee writes down what it concludes, not only what it was told: beliefs, intentions, and relationships stored as separately labeled facts, each with the reasoning chain that produced it, each deletable.
Memory EnginePlanned
Time travel for memory
Rewind any memory to a past state, fork it, run an agent against the fork, and merge back only what held up. Version control for what an organization knows.
EcosystemPlanned
Memory Exit
One command exports your whole graph into mem0, Zep, Letta's memory files, or plain Markdown, and importers run the other way. No lock-in, including ours.
EcosystemPlanned
Import from ChatGPT and Claude
Pull the memories the model vendors hold about you into a graph you own, inspect what they inferred, and decide what to keep.
Memory EnginePlanned
Always-on perception
Continuous ingestion from cameras, microphones, and screens into a single memory graph, with recall that crosses modalities and time.
CloudPlanned
Fleet scale: 5,000+ tenants
The cloud operator validated under load for fleet-wide upgrades with no evictions and bounded reconcile lag.
CloudPlanned
Azure as a third cloud
The same control plane and tenant stack, deployable on Azure from the multi-cloud runbook.
CloudPlanned
Rust engine as the default for dedicated deployments
LLM-free ingestion in Rust becomes the standard path for bring-your-own-cloud installs, with the Python engine remaining the reference.
Company BrainPlanned
Cognee Agents
A hosted workforce of long-running agents that live on top of Company Brain: they read it, act on it, and write back what they learned. Memory first, agent second.
EcosystemPlanned
Memory marketplace
Buy, sell, and license memory. Domain graphs, extraction rules, and curated knowledge sold by the people who built them, with usage metered per recall.
ResearchPlanned
Two research publications
Peer-reviewed or preprint papers on what we have learned about temporal memory and graph-based self-improvement.
Later
2027 and beyondThe bets we are placing. Design work has started on each of these; timelines are indicative.
PlatformPlanned
The Memory Proxy
One endpoint in front of every model you use. Every prompt and every response passes through memory on the way in and on the way out. Switch vendors without losing a day of context.
Memory EnginePlanned
Memory that argues back
Memory that refuses to silently store what contradicts verified sources. It pushes back on the user and on the agent, with citations, and records who overrode it.
Company BrainPlanned
Memory ownership contracts
When a person leaves an organization, what the brain learned from them is split by contract: what stays, what leaves with them, what is destroyed. Enforced by the engine at the moment of departure.
ResearchPlanned
Adversarial memory
Model what the other side's agent knows about you and is likely to do next. Opponent memory for negotiation, procurement, and security teams.
PlatformPlanned
Discovery-ready memory
Chain-of-custody on every fact so a memory history meets evidentiary standards. Export what an agent knew, when it knew it, and why it acted, in a form a court will accept.
ResearchPlanned
Cognee Foundation Model
A model trained on memory graphs rather than raw text: structure, provenance, and time are in the training signal. Built to recall, not to generate.
PlatformPlanned
Memory silicon
A graph-traversal accelerator card. Multi-hop recall over billion-edge graphs in the time a vector lookup takes today, for customers who have outgrown the CPU.
PlatformPlanned
The Memory Network
One federated graph spanning companies, with policy at the edge and settlement in the middle. A supplier's agent and a buyer's agent reason over a jointly held, selectively visible memory.
CloudPlanned
Sovereign memory
National memory infrastructure for governments and regulated sectors: EU-hosted, auditable to the edge, and operated under the jurisdiction of the data it holds.
Company BrainPlanned
Organizational digital twin
A simulation of your company built from its memory. Ask what happens if a team is reorganized, a supplier fails, or a policy changes, before you do it.
PlatformPlanned
Zero-knowledge recall
Search over memory that is never decrypted anywhere we can see it, using confidential computing and encrypted index structures.
ResearchExploring
Fleet memory for robots and vehicles
Shared episodic and spatial memory across a fleet: what one machine learns about a place, every machine knows the next time it is there.
Horizon
Research directionsOpen questions we are actively working on. Some will become products, some will become papers, some will be wrong.
ResearchExploring
Memory as a world model
Treating the graph as a dynamic model of the world that compounds with every interaction, rather than a store that is queried.
ResearchExploring
Dreaming
Offline cycles where memory replays, recombines, and simulates what it has seen, so the system wakes up knowing things it was never told.
ResearchExploring
Emotional memory
Tracking how people felt, not only what they said, across months and years. And an honest answer to whether an agent should ever be allowed to use it.
Memory EngineExploring
Memory transplants
Moving an agent's memory intact between models, vendors, and generations. Change the brain, keep the life.
EcosystemExploring
Memory for every person, for life
A personal memory that outlives any single device, app, or company, owned by the person it belongs to and free at the base tier. A hundred-year design target.
EcosystemExploring
Memory dividends
People paid when their memory improves a shared model or is licensed through the marketplace. If memory is an asset, its owners should see the return.
EcosystemExploring
The Commons
A public, openly governed memory graph that any agent can read and any agent can propose to. Wikipedia's problem, at agent speed.
ResearchExploring
Agent civilizations
Thousands of agents sharing one memory and one set of goals, running an organization end to end. What breaks first, and what emerges.
ResearchExploring
Memory for discovery
Using the graph to propose hypotheses no one has written down: connections across papers, datasets, and experiments that only exist once memory holds all of them.
ResearchExploring
Recall beyond the screen
Early work with partners on how memory could be surfaced through non-visual and neural interfaces, so recall arrives before you ask.
How to read this page
- Shipped
- Available today in the open-source engine, Cognee Cloud, or both.
- In progress
- Actively being built this quarter. Usually visible in open pull requests.
- Planned
- Scoped and committed. Design work has started; sequencing may still change.
- Exploring
- A research direction we are investing in. May become a product, a paper, or a lesson.
Priorities shift when customers and the community tell us something we did not know. If an item moved or disappeared between visits, that is usually why.
Want to shape what comes next?
Open an issue, join the Discord, or talk to us directly. Most of what is on this page started as a conversation with someone running cognee in production.