Store where the memory came from
Mark evidence as user-stated, agent-inferred, or third-party, with confidence, provenance, validity, and action relevance.
source + confidence + provenanceHow it works
Memory should provide evidence. Policy should decide permission.
ContextDB sits between what an agent remembers and what your host lets it do. It evaluates remembered evidence, returns act, ask, or abstain, and stores the reason.
The decision loop
One check before the consequential step.
Ordinary recall can ground a conversation. Action recall is stricter. It only returns evidence that passes the active trust policy.
conversation │ ▼ remember → recall for context │ ├── agent prepares a consequential action ▼ recall_for_action │ ├── act trusted evidence is sufficient ├── ask a specific fact needs attestation └── abstain evidence is missing or policy refuses │ ▼ your host enforces the result → decision record
Apply policy at action time
The same memory can ground conversation while remaining ineligible for a booking, refund, or record update.
evidence × policy → outcomeKeep the reason with the result
Store the outcome, reason, policy version, evidence IDs, partition, and request ID as one decision object.
decision is a first-class objectDecision trace
Answer “why did the agent do that?” without reconstructing a prompt.
A decision record connects the attempted action to the evidence and policy used at that moment. It is designed for debugging and incident review, not as a vague activity log.
action evaluation
act
query
reschedule appointment
reason
trusted user-stated preference
evidence
mem_01JQ6…
policy
trust-v0.3
request
req_8f2…
Core objects
A small model built around decisions.
| Object | What it represents |
|---|---|
| project | One agent deployment with its own keys, policies, and records. |
| user partition | One end-user memory space. The partition key is not authentication. |
| memory | Evidence with source, confidence, provenance, PII annotations, and validity. |
| decision | An act, ask, or abstain outcome stored with its reason and evidence. |
| confirmation | An actor's attestation about a specific fact. Attestation is not objective truth. |
| policy | Versioned rules that determine what evidence is trusted for action. |
Hosted surface
The cloud gateway is intentionally narrow.
The hosted gateway remains Alpha. Routes whose partition ownership cannot yet be proven are blocked, not quietly exposed.
GET /health
GET /ready
POST /v1/remember
POST /v1/remember_many
POST /v1/recall
POST /v1/recall_for_action
GET|POST /v1/pending_confirmations
blocked in hosted alpha
POST /v1/confirm
POST /v1/forget
POST /mcp
Questions
What ContextDB does and does not do.
Does ContextDB execute the business action?
No. ContextDB returns an advisory decision. Your host must enforce that result before it books, refunds, updates, or calls another system.
Does confirmation prove a memory is true?
No. Confirmation records that an actor attested to a specific fact. It does not prove objective truth.
Do I keep my vector store and database?
No. Keep the systems that hold business records and general retrieval. Use ContextDB where remembered evidence can authorize an action.
See the trust model in code.
The SDK and its evals are Apache-2.0 and available now.