Transcript chunks
Text is stored, but meaning is never stabilized.
Governed memory layer for AI
AI memory that actually makes sense
Phren turns conversations into structured, inspectable, reviewable memory so AI can preserve meaning without turning temporary chat into permanent noise.
Early betaFirst 50 testers. Runs on limited infrastructure, so performance may vary.
Works with existing AI models via API, including ChatGPT. Phren adds structured memory and governance on top.
Meaning first
Review before trust
Memory with undo
InputConversation becomes structured candidates.
Memory layer stack
01Chat historyRaw logs
02Session RAMLive context
03Episodic carryoverShort-lived continuity
04Durable memoryGoverned knowledge
OutputReliable context with review, apply, and undo.
4memory layers
9runtime stages
7quality systems
3trust guarantees
Problem
Normal AI memory stores text before meaning.
Without structure, AI memory becomes duplicated fragments, stale assumptions, and untraceable changes.
Duplicate drift
One fact becomes many memories when wording changes.
Permanent noise
Temporary chat can accidentally become long-term context.
Vector-only recall
Similarity finds text, not necessarily trusted knowledge.
No review path
Users cannot inspect, apply, correct, or undo memory changes.
Clarification
Phren is not another chatbot.
Phren does not try to replace AI models. It provides structured memory that works alongside them.
AI modelReasoning, generation, and conversation
PhrenMemory system, governance, and retrieval context
Architecture
Four layers keep temporary context separate from durable knowledge.
Phren does not make every message permanent. It keeps raw logs, live state, short-lived carryover, and governed knowledge in different layers.
A faithful record of what was said. Useful evidence, not automatic truth.
Active anchors, entities, threads, open loops, and recent decisions for the current chat.
What still matters across nearby sessions without storing the whole day again.
Reviewed claims, entities, relations, episodes, and concepts reused later.
Runtime flow
A visible path for every memory decision.
Every turn follows a structured path before anything becomes long-term memory. Retrieval and memory promotion can be inspected instead of guessed.
01chat
02retrieval
03active context
04session RAM
05episodic carryover
06provider reply
07memory processing
08governance
09future retrieval
Capabilities
Memory quality features, grouped like product systems.
Phren combines structure, retrieval, governance, hygiene, and automation boundaries into one operating layer.
Structured objects
Claims, entities, relations, episodes, concepts, lifecycle, and evidence.
Hybrid retrieval
Vectors plus anchors, exact matches, recency, salience, lifecycle, and feedback.
Governed promotion
Candidates become durable through policy, review, or bounded safe automation.
Apply and undo
Memory changes carry diffs, evidence, history, and rollback support.
Memory hygiene
Duplicate pressure, conflicts, canonicalization, and graph health stay visible.
Visible reasoning
Retrieved memory explains why it was selected and what signal made it relevant.
Controlled automation tiers
Low-risk cleanup can run automatically. Identity-risk changes still require review.
Governance and trust
Inspectable, reviewable, reversible.
Memory changes are never hidden. Phren is designed so changes can be understood, corrected, and rolled back when needed.
Inspectable
See why a memory was used, where it came from, and what confidence signals supported it.
Reviewable
Route uncertain memories into inbox, synthesis review, retrieval review, and audits.
Reversible
Keep action history, diff previews, and undo support for memory-changing operations.
Integrations
Works with your existing AI stack.
Phren is an API-first memory layer that can sit behind chat interfaces, provider workflows, and structured retrieval systems.
API-first architecture
Compatible with existing AI providers
Can be used behind chat interfaces
Supports structured retrieval workflows
Simple usage
Use Phren as a normal assistant first.
Users can chat normally. Memory becomes useful over time as Phren tracks active context, proposes candidates, and lets users go deeper only when they need control.
Ask normally
Phren retrieves relevant context
Review what changed
Open graph, inbox, audits, or retrieval review later
Continuous improvement
The memory layer gets sharper without hiding risk.
better retrieval
better continuity
better governance
better compression
better hygiene
Give AI memory a system of record.
Preserve continuity, reduce memory noise, and keep durable knowledge governed.
Early beta: first 50 testersRuns on limited infrastructure. Response speed may vary while the system is optimized for larger workloads.