Available now
Fully supported
Agency
Capture and retrieve governed context directly across Agency sessions.
Test the integrationInstitutional intelligence for people and AI agents
Knowledge informs. Experience teaches. Continuity creates cognition.
Cognisphere connects what your organization learns across apps, agents, and workflows, turning fragmented experiences into reusable intelligence that grows wiser with every interaction.
- 48%
- fewer tokens, one engineer
- 42%
- less model time, one engineer
- 10×
- lower cost per answer, company scale
48% and 42% are measured on one machine across a 100-run benchmark. 10× is the modelled ceiling once the same memory is shared across an organization and every agent surface.
01 / Connected work
Use Cognisphere wherever work happens.
Connect people, agents, apps, and workflows to the same governed learning layer. Start with supported integrations today, with more adapters built on the same interface.
Available now
Available now
Fully supported
VS Code
Capture and retrieve reviewed, evidence-backed knowledge directly within the editor workflow.
Test the integrationRequest project owners for access to Git Repo.
Plug-in roadmap
Next adapters and future surfaces.
ScoutNext
CopilotNext
OpenClawPlanned
AntiGravityPlanned
02 / Benchmark evidence
One engineer. One machine. 48% less context overhead.
A baseline comparison over 100 runs of six questions, with a single engineer reusing only their own accumulated context. Nothing here is shared with a teammate or a second agent yet, which makes this the floor rather than the ceiling.
Token footprint · single seat
48.00%fewer tokens than baseline
92,917,840 tokens avoided over 600 answers — 322,631 per answer at baseline, 167,768 with Cognisphere.
fewer model inferences
36.90% reduction42 seconds saved
42.25% reductionanswers compared
600 per modeBenchmark design
100 runs · 6 questions · baseline vs. Cognisphere
Questions A1–A3 and T1–T3 were evaluated in both modes for every run.
Protect expert time
Reduce repeated questions and context reconstruction so specialists can focus on novel, high-value work.
Preserve institutional intelligence
Carry reviewed decisions, reasoning, and evidence across onboarding, handoffs, teams, and agent sessions.
Make corrections compound
Turn reviewed findings and expert corrections into durable learning instead of one-off answers that disappear.
Results describe this benchmark workload and configuration; production outcomes vary by model, context size, retrieval policy, and task mix.
03 / Scale
One machine proves it. A company multiplies it.
A single engineer can only reuse what they themselves discovered. Share that memory and every discovery is paid for once and recalled everywhere, so answering a question the company has already answered collapses toward the cost of recall.
Model your fleet
VS Code · Agency · PR review · CI · triage · support
- Agent endpoints
- 600
- Questions already answered somewhere
- 97.0%
Modelled cost per answer
10.0×cheaper than re-deriving it
1.9× measured, one seat
10× ceiling
- Tokens avoided per year
- 52.3B
- Tokens per answer
- 32,199
One engineer reuses themselves
You pay to discover something, then spend the rest of the week reusing it. That is the benchmark above: 48% fewer tokens, and the ceiling is roughly 1.9×, because you still had to pay for every discovery yourself.
Every surface shares one memory
VS Code, Agency, PR review bots and CI agents stop keeping private context. A discovery made in one surface is immediately recallable from all of them, so the same spend serves several times the work.
Overlap saturates across a company
Past a few hundred endpoints, almost every question has already been answered by someone. Cost per answer approaches the cost of recall rather than rediscovery — a ceiling of about 10×.
How this is modelled. Cost per answer = overlap × recall + (1 − overlap) × agent run. Overlap rises with reach (engineers × surfaces) and saturates at 97%. The baseline of 322,631 and the agent run of 167,768 tokens per answer are measured in the 100-run benchmark; recall is modelled at 28,000 tokens, and volume assumes 300 agent answers per surface, per engineer, per year. At one engineer on one surface the model returns 1.9×, reproducing the measured result. Everything beyond that point is a projection, not a measurement.
04 / Early feedback
A problem engineering leaders already recognize.
Early reactions highlighted the value of helping people and agents learn from persistent knowledge with the governance required for enterprise use.
Written feedback
Really cool idea and very relevant to where engineering and AI are heading. I like that the team focused not just on agent capabilities, but on one of the biggest challenges today: how to preserve and safely reuse knowledge that would otherwise be lost in individual AI sessions.
The emphasis on governance, security, provenance, and permission-aware access shows strong enterprise thinking. This has the potential to save significant engineering time, improve onboarding and handoffs, and help teams build on each other’s work instead of constantly starting from scratch.
Paraphrased booth feedback
We would use this day to day. Incrementally learning, sharing, and preserving knowledge within and across teams solves a very real engineering need.
Paraphrased booth feedback
A genuinely compelling idea. I am keen to put it to work inside my own team, where I expect it to accelerate initiatives like Tigerlint.
Paraphrased judge feedback
Super impressed by the concept and looking forward to seeing it available alongside today’s Agency and GitHub Copilot capabilities.
Paraphrased booth feedback
This struck a chord. I had already built something along these lines for myself, scoped to a single user rather than a team or an organisation, so my agents could pick up existing context instead of relearning everything from zero. Seeing the same idea applied at org scale is where the real value shows up.
05 / The problem
Every person and agent creates institutional intelligence.
Without a learning layer, the organization keeps paying to rediscover it.
Expertise stays trapped
Valuable decisions, reasoning, and discoveries remain in people’s heads or isolated sessions.
Questions get answered again
People and agents spend time rebuilding context and re-deriving conclusions the organization already earned.
Raw memory creates risk
Unreviewed context can expose data or preserve errors.
06 / How it works
From one-off insight to auditable institutional memory.
Cognisphere distills useful work into structured knowledge, keeps human judgment in control, and returns approved learning when it matters.
Opted-in session
Capture
Record bounded human and agent work without running inference in the capture path.
Local short-term memory
Distill
Convert useful events into compact, structured summaries, entities, relationships, and supporting evidence.
Human checkpoint
Review
A reviewer inspects the exact candidate graph, provenance, branch, and supporting evidence if flagged by pipeline.
Governed shared knowledge
Reuse
Return approved context to authorized people and agents, then retract it when needed.
07 / Cross-session journey
One session learns. The organization remembers.
Run the story to see one solved problem become durable institutional advantage for whoever works on it next.
Agent session A
Investigation
✓
Evidence-backed fix discovered
1
Capture
2
Screen
3
Review
4
Promote
5
Recall
Agent session B
Future session
Governed context
Retry policy → fixed by → bounded timeout
Approved evidence · Authorized retrieval08 / Guardrails by design
Trusted memory needs deliberate boundaries.
Cognisphere does not treat every model output as organizational truth. Knowledge remains a candidate until its evidence and provenance pass a deliberate promotion boundary.
Explicit participation
Memory capture is designed to begin with deliberate user and workload participation.
Content safety
Sensitive-content checks and policy screening reduce risk before knowledge is persisted.
Human validation
High-value knowledge crosses an explicit review boundary before permanent promotion.
Permission-aware access
Read, write, and promotion policies control how shared context is used.
Traceable and retractable
Provenance supports auditability, correction, and removal throughout the memory lifecycle.
09 / Roadmap
From governed memory to an enterprise-ready context fabric.
The roadmap advances Cognisphere from a connected agent experience toward cloud scale, responsible AI integration, and first-party distribution.
Now
Connected experience
Agent memory foundation
- Cross-session context
- Human validation and provenance
- Permissions and lifecycle guardrails
Beta · In progress
Cloud-based
Shared governed context
- Cloud-hosted memory and scalable retrieval
- Stronger identity boundaries
- Managed agent integrations
- Enhanced collaboration tools (Scout, AntiGravity, Openclaw, Copilot)
GA readiness
Enterprise distribution
Responsible and secure by default
- Responsible AI integration
- Microsoft Defender security scanning
- Marketplace first-party application
Compound what your organization learns
