Capture, structure, and recall the right context across every agent and workflow.
Personize turns scattered customer history into a persistent, governed memory layer so AI agents stay accurate, consistent, and useful across time and tools.
Enterprise-grade accuracy
AI memory fails when it is inconsistent, unstructured, or not grounded in record-level truth. Personize is designed to keep memory accurate across extraction, recall, and ongoing updates.
Memorization
Extract every insight that exists in your data—nothing skipped, nothing lost.
Recall
Surface the right memories at the right time for each task.
Governance
Built-in PII redaction, tiered data access, and full audit trails for compliance.
Record Consistency
Same customer = same facts, no matter how many memories exist.
Agent Consistency
Every agent sees the same truth, every time, across every task.
Deduplication
Same insight, one record — no matter how many sources mention it.
Structured for automation. Rich for agent reasoning.
Most customer insights belong in structured fields that workflows can act on. The rest belongs in high-signal context that improves relevance.
Schema-Enforced
- •Schema-enforced attributes with types and allowed values
- •Queryable fields for segmentation and filtering
- •Predictable formats for workflow automation
Free-Form
- •Details that don't fit a schema
- •High-signal context that improves relevance
- •The human stuff AI usually forgets
You could write the SQL. Or make one call.
SQL answers “which rows match these conditions?” Retrieval answers “what’s most relevant to know about this entity, right now, ranked and ready for an agent?” They’re different jobs — and you get both.
Raw SQL you build & maintain
To turn “what should I know before this renewal call?” into context, you’d write and own:
- •Embed the query, then vector-search with the right model and dimension
- •Blend semantic + keyword (BM25) + recency, then re-rank
- •Resolve identity yourself (which record is this email or domain?)
- •Join memories, governance docs, and the graph into one result
- •Re-apply scoping, redaction, soft-delete, and staleness on every query
- •Keep it all correct as the schema and ranking evolve
One retrieve call
POST /memory/retrieve
{ "email": "sarah@meridian.com",
"query": "renewal risk",
"mode": "brief" }- Relevance, not row-matching — hybrid semantic + keyword + recency, reranked
- Identity resolution built in — ask by email or domain
- Properties + memories + related documents + graph, fused in one result
- Governance applied automatically — scoping, redaction, soft-delete, staleness
- Modes for the job: scout, brief (LLM answer + citations), filter, expand, fetch
- Schema-abstracted — it gets better as our ranking improves, no query rewrites
Need exact rows? Filter mode compiles to SQL for you.
For deterministic queries — “every healthcare contact with budget over $100k and status active” — call retrieve in filter mode. You pass structured filters; we compile them to SQL against your Postgres and apply the same governance. You get SQL’s precision without writing or maintaining SQL — and semantic relevance in the same API when you need it. You never have to choose.
It’s your Postgres — run SQL for analytics and reporting all you like. Unified retrieval is the part you’d otherwise build: a governed, semantic, multi-source engine that hands an agent the right context in one call.
Memory Quality — Where We Stand
We combine an industry-standard benchmark with proprietary experiments to prove — not claim — that our memory works.
LoCoMo Benchmark
LoCoMo (Snap Research, ACL 2024) is the industry's most rigorous benchmark for long-term conversational memory. It tests whether a system can recall specific facts, connect information across separate conversations, understand temporal order, reason beyond what was explicitly said, and know when to say "I don't know" — the way a trusted colleague would over dozens of sessions.
Our overall accuracy is ~75% and improving. We already exceed human-level performance on open-ended inference. Comparable systems score 42–67%.
A detailed paper with final results will accompany our formal launch.
LoCoMo Overall Accuracy
What Only We Measure
Beyond LoCoMo — 13 controlled experiments with no industry equivalent.
Fact Extraction
Across call notes, documents, chat, transcripts, and email.
Combined Recall
Structured + unstructured memory working together.
Cross-Entity Leak
Client A's data never appears in Client B's context.
Governance Routing
Right rules applied to the right situation, every time.
Token Savings
Progressive context reuse in multi-step workflows.
Deduplication
Memory stays lean and accurate as data grows.
Memory Economics
See the difference structured, compressed, intelligently retrieved memory makes to your AI costs.
Traditional Approach
Every interaction re-processes full unstructured context—verbose, redundant, and expensive.
Costs scale linearly with every interaction
Personize Memory
Structured, compressed memories retrieved adaptively—only relevant context, in the right format, at the right depth.
Costs stay flat as knowledge deepens
Dramatically Lower Costs
with higher output quality—because less noise means better answers
Optimized recall, not just storage
Reflection-expanded recall
Personize expands a simple question into a complete, expert-level retrieval plan so agents do not miss what matters.
Guided usage
Return the relevant policies and best practices alongside memory so agents know how to apply context correctly.
Continuous quality
Monitor memory accuracy, detect drift, and tune extraction rules — before it impacts customers.
Delivered to any agent
MCP and API delivery so any AI tool or framework can use the same customer memory.
Turn customer history into action
Capture insights once. Use them across every GTM workflow.