Learn from every source
Extract memory from free text, write exact structured records, import batches, and preserve long-form company documents.
Personize Private
Memory Box.Your Enterprise Data as private memory, accessible to all your AI agents on any platform, inside your walls.
Local. Private. Inspectable. Built for the memory your business cannot afford to lose.
Memory BoxInside your infrastructure
Runs on your hardware
Uses your PostgreSQL
Local models per function
Personize stays out of the data path
Your company already knows
Put its memory to work. Keep it private.
Give agents durable context without handing your institutional memory to another vendor. Memory Box turns the knowledge already inside your business into governed, reusable AI memory.
Available in 0.6.6
While your team sleeps, Memory Box improves what it knows.
Dreaming
Dreaming is an idle-gated AI pass that reviews selected company records, finds memory problems, and consolidates the best version of what your business knows.
It works against your PostgreSQL and your chosen model. Point it at Ollama or vLLM for a fully local overnight pass.
Inspect. Compare. Resolve. Improve.
01Dedupe duplicate facts
02Resolve contradictions
03Merge scattered profiles
04Retire stale values
05Promote recurring insights
06Distill long documents
Idle-gated
Yields to live work
Dry-run mode
Preview before applying
History preserved
Nothing hard-deleted
Every run audited
Review what changed
A complete memory operating layer
It does more than store what happened.
Memory Box captures company knowledge, retrieves the right context, applies your rules, connects relationships, and gives every agent one governed memory to work from.
Retrieve the right depth
Move from a fast scout to a synthesized brief, deep session context, structured filters, or direct record fetches.
Map company relationships
Infer and store relationships across contacts, companies, deals, repositories, incidents, or any entities you define.
Apply governance at the moment of work
Store guidelines, playbooks, references, and SOPs, then route the most relevant rules into every AI task.
Generate with memory
Run governed prompts with retrieved memory and policies already attached, then optionally memorize useful output.
Model your business, not ours
Define custom entities, collections, properties, relationships, document types, and complete reusable schema kits.
Connect every agent
Use the bundled MCP server, REST API, or TypeScript SDK so Claude, Cursor, and your own agents share one memory.
Keep background work durable
Queue bulk imports and long-running prompts in PostgreSQL with bounded concurrency, timeouts, and crash recovery.
Choose where every function runs
Fully local when you need it. Protected when you reach out.
Memory Box gives you a deliberate choice per function. Keep everything local, or use an approved external model without exposing raw secrets and identifiers.
Mode 01
Fully local
Point inference and embeddings at local models. Stored text, queries, retrieval, and generation stay on your infrastructure.
Zero memory-content egress
Mode 02
Protected external models
Strip secrets and financial identifiers. Swap known personal identifiers for realistic stand-ins. Restore the real values on your side.
Real identities stay on your box
Private by architecture
Security controls your team can inspect.
Privacy is not a promise layered over someone else's data path. It is built into where Memory Box runs, what it stores, and how every external call is protected.
Local-first by function
Run extraction, generation, retrieval synthesis, intent, and embeddings on local models when you need zero content egress.
Redaction before egress
Secrets, card numbers, IBANs, and SSNs are removed before any approved external model call.
Identifiers become stand-ins
Emails, phones, IPs, and your known names and companies are replaced, then restored locally after the model responds.
Your patterns, your rules
Add protection for account numbers, policy IDs, claim numbers, medical record numbers, or identifiers unique to your business.
Inspectable privacy layer
Use the built-in detector, extend it, or replace it. Enforcement and local restoration remain part of the product.
Fails toward privacy
If protection errors, Memory Box applies a safe fallback or refuses the external call. Raw text is not the fallback.
hash + counts, never content
Provenance without exposure
Prove what left the box without storing what left.
Turn on the provenance log for a per-call record of the protection checkpoint, protected entity types and counts, and a hash of the outbound payload. The content itself is never logged.
Protection checkpoint
Entity types and counts
Outbound payload hash
No content stored in the log
One container. Your environment.
Your hardware. Your database. Your model choices.
Deploy Memory Box on infrastructure you control, connect your PostgreSQL, and choose local or approved external models per function. Personize receives usage counts, never memory content.
Run the container
Deploy inside your controlled environment.
Connect your PostgreSQL
Keep durable memory in the database you own.
Choose each model path
Route every function locally or through protection.
Questions security teams ask
Clear answers before you bring memory inside.
Does data ever leave Memory Box?
You choose. Run inference and embeddings locally and no memory content leaves your network. If you use an approved external model, sensitive fields are removed or replaced before the call and restored locally afterward.
Where is memory stored?
Memory Box runs against PostgreSQL that you control. Personize is not in the memory-content data path.
Can we protect our own identifiers?
Yes. Add patterns for the identifiers specific to your business, such as account, policy, claim, case, or medical record numbers, without forking the product.
Can security teams inspect what leaves?
Yes. The optional provenance log records the protection checkpoint, protected entity types and counts, and a hash of what was sent. It does not store the content itself.
What does Dreaming change?
Dreaming can correct property values while preserving prior values in history, and it can update long-form documents in place. Nothing is hard-deleted. Use dry-run mode to preview proposed changes before applying them.
Can Dreaming run fully locally?
Yes. Point Dreaming at a local runtime such as Ollama or vLLM for a fully private overnight consolidation pass against your PostgreSQL.
Personize Private Memory Box
Memory compounds. So does the advantage.
Put private, governed AI memory inside your walls and make the context your company already owns work harder.
Talk to us about Memory Box