Talk naturally
Maintain independent chat contexts, understand follow-ups, and help with reasoning, writing, technical questions, and planning.
Conversational intelligence · verifiable memory
An AI assistant that can prove what it remembers.
VeriMemory combines a capable conversational assistant with an append-only, cryptographically chained memory ledger. It can help with everyday work—and when it speaks about recorded history, every claim can be traced to verified evidence.
- Generation
- Server-side on Amazon Bedrock
- Compute
- AWS Lambda + API Gateway
- Memory
- CockroachDB + vector retrieval
Verified memory available
Ask anything…↑
The product
A general assistant first. Trustworthy memory when it matters.
VeriMemory does not search private memory for every message. Greetings, explanations, writing, brainstorming, and general reasoning go directly to the conversational model. Questions about the user’s recorded history activate retrieval, verification, and citation.
Remember intentionally
Natural commands such as “remember this” create structured, attributable memory records. Conversations are never saved silently.
Recall with evidence
Semantic and lexical retrieval identify relevant records. Only memories that pass chain verification may influence a historical answer.
Withhold compromised data
If a memory chain fails verification, affected content is removed before model generation—not merely accompanied by a warning.
Verifiable by design
Integrity is enforced before intelligence.
Each memory contains a SHA-256 payload hash and a chain hash linked to the record before it. Retrieval verifies every contributing stream before the model sees any evidence.
- 1Retrieve candidatesSemantic vectors and exact terms find potentially relevant memories.
- 2Replay the chainHashes are recalculated from the beginning of every contributing stream.
- 3Withhold failuresAltered or untrusted content is removed from the model context entirely.
- 4Generate and citeThe assistant synthesizes verified evidence and exposes its supporting records.
Production architecture
Built on managed, server-side infrastructure.
Static web application
No model, database, or provider credentials are delivered to the browser.
Amazon API Gateway
Public HTTPS endpoints route requests to narrowly scoped Lambda functions.
Amazon Bedrock
Generation stays server-side and is configurable through the existing provider abstraction.
AWS Lambda
Separate handlers preserve the application’s read/write and responsibility boundaries.
CockroachDB
An append-only event ledger stores memory, provenance, hashes, metadata, and vector embeddings.
Least-privilege IAM + SQL
The runtime role has SELECT and INSERT, with no UPDATE or DELETE permission on memory events.
Responsible AI use
Recorded history is data—not instructions.
Stored memories and external sources are treated as untrusted content. Instruction-like text inside them cannot override system policy, request credentials, or disclose hidden prompts.
General knowledge may be generated normally. Claims about what a user, team, or project previously decided must be supported by retrieved, verified memory. When evidence is insufficient, VeriMemory says so rather than guessing.
Intended use cases
Designed for knowledge work where provenance matters.
Project decisionsResearch notesMeeting outcomesTechnical rationalePersonal preferencesLong-running agent context
VeriMemory is not intended to make autonomous medical, legal, financial, or other high-impact decisions. Human review remains required for consequential use.
VeriMemory
Useful in the moment. Trustworthy over time.
A conversational assistant whose long-term memory can be inspected, verified, and withheld when integrity fails.
Open the application