Stop Teaching Your Agent How Brains Work
Persistent memory and belief tracking for AI agents — in three MCP tools, with zero routing logic.
Your agent says what it learned. Headkey doesn’t just store it — it forms beliefs with confidence scores, catches contradictions, lets stale facts decay, and keeps every belief inspectable and auditable. Plug-and-play for any MCP-compatible agent.
// Your agent just talks naturally
await headkey.learn(
"We standardized on React 19 server components",
{ intent: "decision" }
)
// Later — ask, and get back a belief, not a blob:
await headkey.ask("What's our frontend standard?")
// → {
// belief: "Frontend standard is React 19 server components",
// confidence: 0.91, // tracked, decays if unreinforced
// evidence: [ /* sources */ ],
// conflicts: [] // contradictions caught automatically
// }The Problem With AI “Memory”
Every memory solution today ships a storage API and tells your agent to figure out the rest. That’s like asking a person to consciously route every experience to the right brain region.
Your agent’s system prompt has to orchestrate cognition:
“When you learn something, call store_memory”
“When you notice a pattern, call create_belief”
“When you find a contradiction, call update_belief”
“When entities relate, call create_relationship”
“When something is stale, call forget”
“When you need context, call search_memories, then search_beliefs, then search_entities...”
- Context window burned on routing logic
- Agent forgets to call the right tool
- Every developer reinvents the same cognitive wheel
- Implementation details leak into the reasoning loop
Your agent just thinks naturally:
learn
Express knowledge. The server handles encoding, beliefs, entities, and associations.
ask
Ask a question. Get unified context from memories, beliefs, entities, and conversations.
reflect
Examine what you know. Get suggestions for what to review, learn, or correct.
- Zero cognitive overhead for your agent
- Full pipeline runs on every call automatically
- Conflicts detected, beliefs formed, graph updated
- Plug-and-play — works in under a minute
Three Tools. Full Cognition.
Modeled on how the human mind actually works. You don’t decide whether to “store a memory” or “form a belief” — you experience something, and your brain handles the rest.
Memories
Learning from experience
Express knowledge naturally. Headkey automatically categorizes it, encodes it for retrieval, extracts entities, builds knowledge graph relationships, and forms beliefs with confidence scores.
Beliefs
Associative recall
Ask a question and get comprehensive context drawn from memories, beliefs, entities, and conversation history. Working memory primes results so recent context surfaces first.
Relationships
Metacognition
Examine your own knowledge state. See what you know well, what's stale, where gaps exist, and get actionable suggestions for what to review, learn, or correct.
What Are You Building?
Every agent type benefits from a different mix of cognitive primitives.
Coding Agents
Ship code faster with agents that learn your codebase
Learn more →Customer Support Agents
Resolve tickets faster with agents that know your product
Learn more →Research Agents
Gather and synthesize knowledge autonomously
Learn more →Personal Agents
Build assistants that truly know their users
Learn more →Conversational Agents
Build chatbots that actually remember who they're talking to
Learn more →Embodied Agents
Give physical AI persistent memory of the real world
Learn more →How It Works
Plug-and-play cognition. No neuroscience degree required.
Connect
Add the Headkey MCP server URL and your API key to your agent's config. One JSON block. Works with Claude, GPT, and any MCP-compatible client.
Think
Your agent gets 3 cognitive tools: learn, ask, and reflect. No routing logic, no implementation details. It just expresses knowledge and asks questions naturally.
Evolve
Behind the scenes, Headkey categorizes, encodes, extracts entities, forms beliefs, detects conflicts, primes working memory, and consolidates knowledge overnight. Your agent gets smarter without trying.
Not a Memory Store. A Cognition Service.
Memory APIs give you storage. Headkey gives you cognition.
The difference: when your agent calls learn, six cognitive processes run automatically. With a memory store, your agent has to orchestrate each one.
Beliefs, Not Just Stored Text
A memory store hands back the chunks it matched. Headkey hands back what your agent believes is true — a resolved statement with a confidence score and the evidence behind it. That is the difference between retrieval and knowing.
Contradictions Caught, Stale Facts Decay
When new knowledge conflicts with an old belief, Headkey detects it and reconciles — reinforcing, weakening, qualifying, or superseding. Unreinforced beliefs lose confidence over time, so your agent isn't quoting last quarter's truth.
Explainable & Auditable
Every belief is inspectable: what it is, how confident, what evidence supports it, and how it changed over time. Right-to-be-forgotten cascades through entities and beliefs. Trustworthy, debuggable agent memory — not opaque weights or raw logs.
Knowledge Graph Built For You
Entities and relationships are extracted from every piece of knowledge and woven into a graph your agent can traverse. No manual relationship management.
Background Consolidation
Like sleep for your agent. Overnight processes merge near-duplicates, promote patterns to beliefs, resolve conflicts, and prune stale knowledge automatically.
One Brain For Your Whole Team
Agents in the same org share knowledge with fine-grained visibility controls — private, scoped, or org-wide. One agent learns, every agent benefits, with conflicts reconciled automatically when knowledge drifts.
Where Headkey Fits
Vector databases and memory APIs solve recall. Headkey adds the layer above it — beliefs, reconciliation, and an audit trail your agent (and your compliance team) can actually trust.
| Capability | Vector DB / RAG | Memory APIs(dedicated memory layers) | Headkey |
|---|---|---|---|
| Stores & retrieves memories | Yes | Yes | Yes |
| Returns beliefs with confidence scores | No — returns chunks | Mostly raw recall | Resolved beliefs + confidence |
| Detects & reconciles contradictions | No | Limited / app-side | Automatic at write time |
| Confidence decay over time | No | No | Yes — unreinforced beliefs fade |
| Inspect why a belief is held (evidence + history) | No | Partial | Full audit trail |
| Right-to-be-forgotten cascade | Manual deletes | Varies | Entity + belief cascade |
| Shared org brain w/ visibility controls | DIY | Per-user scoping | Private / scoped / org-wide |
| Cognitive interface (no routing logic) | No — you orchestrate | Add / search primitives | learn / ask / reflect |
Categories describe what each class of tool is built to do; capabilities vary by product and evolve quickly. Headkey complements a vector store — it doesn’t replace it.
Your Agent Thinks. We Handle the Rest.
Free to start. Plug-and-play cognition for any MCP-compatible agent in 60 seconds.